pdfimage.py

#!/usr/bin/env python3

# Write this man about the bug in PIL Image.paste()!
# https://sudonull.com/post/129265-Alpha_composite-optimization-history-in-Pillow-20
# https://habr.com/ru/post/98743/
# https:/homm/
# https://twitter.com/wouldntfix

from pdfrw import PdfReader, PdfWriter, PdfObject, PdfName, PdfArray, PdfDict, IndirectPdfDict
from pdfrw import py23_diffs # Ugly, but necessary: https:/pmaupin/pdfrw/issues/161

from pdfrwx.common import err, msg, warn, eprint, get_key, get_any_key
from pdfrwx.pdffilter import PdfFilter
from pdfrwx.pdfobjects import PdfObjects
from pdfrwx.pdffunctionparser import PdfFunction

from simage import SImage

from attrdict.attrdict import AttrDict

from typing import Union

import argparse, struct, zlib, base64, sys, re, os, subprocess, tempfile

from io import BytesIO
import numpy as np
import shutil

from PIL import Image, TiffImagePlugin, ImageChops, ImageCms
Image.MAX_IMAGE_PIXELS = None

from glymur import Jp2k

from pprint import pprint

CMYK_DEFAULT_ICC_PROFILE = open(os.path.join(os.path.dirname(__file__),'color_profiles/USWebCoatedSWOP.icc'), 'rb').read()

# ============================================================== timeit

from time import time
tStart = time()
def timeit(s): print(f'TIME: {s}: {time()-tStart:0.3f} sec')

# ============================================================== class OS

class OS:

    def execute(cmd:list[str]):
        '''
        Execute command + arguments given a list: cmd[0] is the executable's name,
        cmd[1:] are the command's arguments. On Windows, the .exe extension should be omitted.
        '''
        exe = '.exe' if sys.platform == 'win32' else ''
        result = subprocess.run([cmd[0] + exe] + cmd[1:], capture_output=True)
        if result.returncode != 0: sys.exit(f'execute() error: {result.stderr.decode()}')
        return result.stdout.decode()

# ========================================================================== class PdfColorSpace

# Typedef
    
CS_TYPE = Union[PdfName, PdfArray]

class PdfColorSpace:

    # A mapping from color space name to components per pixel; None means that cpp varies
    spaces = {
        '/DeviceGray': 1, '/CalGray': 1, '/Indexed': 1, '/Separation': 1,
        '/DeviceRGB': 3, '/CalRGB': 3, '/Lab': 3, '/DeviceCMYK': 4,
        '/ICCBased': None, '/DeviceN': None, '/NChannel': None
    }

    # A mapping from components per pixel to PIL image modes
    modes = {1:'L', 3:'RGB', 4:'CMYK'}

    @staticmethod
    def toStr(cs:CS_TYPE):
        return cs if isinstance(cs, str) \
            else [type(x) if type(x) in [PdfArray, PdfDict, bytes] else x for x in cs]

    @staticmethod
    def get_name(cs:CS_TYPE):
        '''
        Returns the name of the color space (on of the PdfColorSpace.spaces.keys())
        '''
        name = cs if not isinstance(cs,PdfArray) \
                    else '/NChannel' if cs[0] == '/DeviceN' and len(cs) == 5 and cs[4].Subtype == '/NChannel' \
                    else cs[0]
        if name != None and name not in PdfColorSpace.spaces:
            raise ValueError(f'bad color space: {cs}')
        return name

    @staticmethod
    def get_cpp(cs:CS_TYPE):
        '''
        Return the number of components per pixel for a give color space.
        '''
        name = PdfColorSpace.get_name(cs)
        return int(cs[1].N) if name == '/ICCBased' else \
                len(cs[1]) if name in ['/DeviceN', '/NChannel'] else \
                PdfColorSpace.spaces.get(name)

    @staticmethod
    def get_mode(cs:CS_TYPE, bpc:int):
        '''
        '''
        return '1' if bpc == 1 else PdfColorSpace.modes.get(PdfColorSpace.get_cpp(cs))

    @staticmethod
    def get_icc_profile(cs:CS_TYPE):
        '''
        For calibrated /CalGray, /CalRGB, /Lab and /ICCBased color spaces, returns
        the ICC color profile as a bytes object. For the /Indexed color space,
        returns the ICC color profile of the 'base' color space, if present, or None.
        For all other color spaces returns None.
        '''
        name = PdfColorSpace.get_name(cs)
        if name == '/Indexed': return PdfColorSpace.get_icc_profile(cs[1])
        elif name == '/ICCBased': return py23_diffs.convert_store(PdfFilter.uncompress(cs[1]).stream)
        elif name in ['/CalGray','/CalRGB', '/Lab']: return PdfColorSpace.create_profile(cs)
        else: None

    @staticmethod
    def get_palette(cs:CS_TYPE):
        '''
        Return a tuple (paletteColorSpace, paletteArray), where paletteArray is a numpy array of shape (-1, cpp),
        cpp is the number of color components in the paletteColorSpace.
        '''
        if PdfColorSpace.get_name(cs) != '/Indexed': return None, None
        name, base, hival, pal = cs

        # pal is either a IndirectPdfDict, with palette in its stream, or a PdfString
        pal = py23_diffs.convert_store(PdfFilter.uncompress(pal).stream) if isinstance(pal,PdfDict) \
                else pal.to_bytes()

        # Fix the tail if necessary
        palette_cpp = PdfColorSpace.get_cpp(base)
        size = (int(hival) + 1) * palette_cpp
        while len(pal) > size and pal[-1] in b'\n\r':
            pal = pal[:-1]
        if len(pal) != size:
            raise ValueError(f'palette size mismatch: expected {size}, got {len(pal)}')

        return base, np.frombuffer(pal, 'uint8').reshape(-1, palette_cpp)

    @staticmethod
    def create_profile(cs:CS_TYPE):
        '''
        Creates an ICC profile for a given color space, when it's one of ['/CalGray', ..], ['/CalRGB', ..], ['/Lab', ..]
        '''
        name = PdfColorSpace.get_name(cs)
        if name not in ['/CalGray','/CalRGB', '/Lab']: return None
        dic = cs[1]

        # Calculate the CIR xyY values of the white point
        X,Y,Z = [float(v) for v in dic.WhitePoint]
        x,y = X/(X+Y+Z), Y/(X+Y+Z)
        try:
            # LittleCMS allows to specify the white point directly when creating a Lab profile
            import littlecms as lc
            ctx = lc.cmsCreateContext(None, None)
            white = lc.cmsCIExyY()
            white.x, white.y, white.Y = x, y, Y
            if name == '/Lab':
                profile = lc.cmsCreateLab2Profile(white)
            if name == '/CalGray':
                gamma = float(dic.Gamma) if dic.Gamma != None else 1
                transferFunction = lc.cmsBuildGamma(ctx,gamma)
                profile = lc.cmsCreateGrayProfile(white, transferFunction)
            if name == '/CalRGB':
                primaries = [float(v) for v in dic.Matrix] if dic.Matrix != None else [1,0,0,0,1,0,0,0,1]
                primaries = [primaries[3*i,3*i+3] for i in range(3)]
                primaries = [[X/(X+Y+Z),Y/(X+Y+Z),Y] for X,Y,Z in primaries]
                gamma = [float(v) for v in dic.Gamma] if dic.Gamma != None else [1,1,1]
                transferFunction = [lc.cmsBuildGamma(ctx, g) for g in gamma]
                profile = lc.cmsCreateRGBProfile(white, primaries, transferFunction)
            with BytesIO() as bs:
                lc.cmsSaveProfileToStream(profile, bs)
                icc_profile = bs.getvalue()
        except:
            warn('LittleCMS not installed')
            warn('using ImageCMS with the (approximate) McCamy\'s formula for correlated temperature')
            n = (x - 0.3320) / (0.1858 - y)
            CCT = 449 * n*n*n + 3525 * n*n + 6823.3 * n + 5520.33
            if name == '/Lab':
                icc_profile = ImageCms.createProfile('LAB', colorTemp = CCT)
            if name == '/CalRGB':
                return None

        if dic.BlackPoint != None: warn(f'/BlackPoint ignored: {cs}')
        if dic.Range != None: warn(f'/Range ignored: {cs}')

        return icc_profile
    
    @staticmethod
    def reduce(cs:CS_TYPE, array:np.ndarray, Decode:list[float] = None, bpc:int = 8, mask:PdfDict = None):
        '''
        Reduces the image array:
        
        * decodes it using the specified Decode array if it's not None or otherwise
        the default decode array, which is determined based on the specified colorspace and bpc;
        * un-multiplies alpha if mask.Matte is not None;
        * if the colorspace is one of `/Separation`, `/DeviceN`, `/NChannel`, reduces the colorspace
        by remapping the image array to the corresponding alternate colorspace;
        * encodes the image array using the default Decode array based on the ending colorspace
        (the original or the alternate one, depending on whether the colorspace has been reduced
        in the previous step) and the value of bpc == 8.
        
        Returns the tuple (newColorSpace, array).
        '''
        SPECIAL = ['/Separation', '/DeviceN', '/NChannel']

        name = PdfColorSpace.get_name(cs)
        h,w = array.shape[:2]
        assert array.ndim in [2,3]

        if bpc == 1:
            warn(f'cannot apply colorspace {name} to a bitonal image; skipping')
            return cs, array

        msg(f'applying color space: {name}')
        Matte = mask.Matte if mask else None
        default_decode = PdfDecodeArray.get_default(cs, bpc)
        applyDecode = bpc != 8 or Matte or Decode and ((name in SPECIAL) or (Decode != default_decode))

        # Decode
        if applyDecode:
            Decode = Decode or default_decode
            array = PdfDecodeArray.decode(array, Decode, bpc)

        # Matte
        if Matte:

            # Get matte array
            matte_array = np.array([float(x) for x in Matte])[np.newaxis, np.newaxis, :]

            # Render alpha
            alpha = PdfImage(obj = mask)
            alpha.render()

            # Resize alpha
            if alpha.w() != w or alpha.h() != h:
                msg(f'resizing alpha: {alpha.w()} x {alpha.h()} --> {w} x {h}')
                alpha.resize(w,h)

            # Get decoded alpha array
            alpha_array = alpha.get_array()
            alpha_decode_default = PdfDecodeArray.get_default(alpha.ColorSpace, alpha.bpc)
            alpha_array = PdfDecodeArray.decode(alpha.get_array(), alpha_decode_default, alpha.bpc)
            alpha_array = alpha_array[:,:,np.newaxis]

            # Un-multiply alpha
            msg(f'Unmultiplying alpha; Matte = {Matte}')
            array =  matte_array + (array - matte_array) / np.maximum(alpha_array, 0.000001)

            # Remove Matte entry from the mask
            mask.Matte = None

        # These color spaces are function-based, so apply this function
        if name in SPECIAL:

            altColorSpace, colorTransformFunction = cs[2], cs[3]
            array = array if array.ndim == 3 else np.dstack([array])
            stack = np.moveaxis(array,-1,0)
            stack = PdfFunction(colorTransformFunction).process(stack)
            array = np.moveaxis(stack,0,-1)
            cs = altColorSpace

        # Encode with bpc == 8
        if applyDecode:
            array = PdfDecodeArray.encode(array, PdfDecodeArray.get_default(cs, 8))

        return cs, array


    @staticmethod
    def apply_default_page_colorspace_icc_profile(page:PdfDict, cs:CS_TYPE):
        '''
        If image.inf['icc_profile'] == None, apply the ICC profile from the page's default color space
        (an entry in page.Resources.Colorspace, if present) to the image (in-place).
        '''
        if page == None or image.info.get('icc_profile'): return
        try:
            default_cs = page.Resources.ColorSpace[cs]
        except:
            pass

    # @staticmethod
    # def make_indexed_colorspace(palette:bytes, baseColorspace):
    #     '''
    #     Returns an /Indexed colorspace based on the image's palette and baseColorspace, or
    #     None if image has no palette.
    #     '''
    #     # palette = image.getpalette()
    #     # if palette == None: return None
    #     # palette = b''.join(v.to_bytes(1,'big') for v in image.getpalette())

    #     cpp = PdfColorSpace.get_cpp(baseColorspace)
    #     assert len(palette) % cpp == 0
        
    #     palette_xobj = IndirectPdfDict(Filter = PdfName.FlateDecode,
    #                                    stream = zlib.compress(palette).decode('Latin-1'))
    #     return PdfArray([PdfName.Indexed, baseColorspace, len(palette) // cpp, palette_xobj])

    @staticmethod
    def make_icc_based_colorspace(icc_profile:bytes, N:int):
        '''
        Returns an /ICCBased colorspace made from the icc_profile with N being the number of color components.
        '''
        icc_xobj = IndirectPdfDict(
            N = N,
            Filter = PdfName.FlateDecode,
            stream = zlib.compress(icc_profile).decode('Latin-1')
        )
        return PdfArray([PdfName.ICCBased, icc_xobj])
    
# ========================================================================== class PdfDecodeArray

class PdfDecodeArray:

    '''
    The class facilitates processing of image.Decode arrays.
    '''

    @staticmethod
    def get_actual(Decode:PdfArray, cs:CS_TYPE, bpc:int):
        '''
        Returns actual Decode array as list[float] if self.image.Decode != None, or self.get_default(colorSpace) otherwise.
        ''' 
        FLOAT = lambda array: [float(a) for a in array]
        return FLOAT(Decode) if Decode != None else PdfDecodeArray.get_default(cs, bpc)

    @staticmethod
    def get_default(cs:CS_TYPE, bpc:int):
        '''
        Returns the default Decode array for a give color space; see Adobe PDF Ref. 1.7, table 4.40 (p.345)
        ''' 
        FLOAT = lambda array: [float(a) for a in array]

        name = PdfColorSpace.get_name(cs)
        cpp = PdfColorSpace.get_cpp(cs)

        if name == '/Lab':
            d = [0.0, 100.0] + (FLOAT(cs[1].Range) if cs[1].Range != None else [-100.0, 100.0] * 2)
        elif name == '/ICCBased' and cs[1].Range != None:
            d = FLOAT(cs[1].Range)
        elif name == '/Indexed':
            d = [0, (1<<bpc)-1]
        elif name == '/Pattern':
            d = None
        else:
            d = [0.0, 1.0] * cpp

        return d

    @staticmethod
    def decode(array:np.ndarray, Decode:list[float], bpc:int):
        '''
        Translates a numpy array from encoded ("stream space": int, 1..2**bpc -1)
        to decoded ("image space", float: mostly 0.0..1.0) representation.
        The array should be (D+1) dimensional, where D is the # of dimensions of the image
        and the last dimension is for the color index.
        '''
        assert Decode
        msg(f'applying Decode: {Decode}')
        INTERPOLATE = lambda x, xmin, xmax, ymin, ymax: ymin + ((x - xmin) * (ymax - ymin)) / (xmax - xmin)
        iMax = float((1<<bpc) - 1)
        if array.ndim == 2: array = np.dstack([array])
        N = array.shape[-1] # number of colors
        assert len(Decode) == 2*N
        array = np.clip(array, 0, iMax).astype(float)
        for i in range(N):
            array[...,i] = INTERPOLATE(array[...,i], 0.0, iMax, Decode[2*i], Decode[2*i + 1])
        return array if N > 1 else array[...,0]

    @staticmethod
    def encode(array:np.ndarray, Decode:list[float]):
        '''
        Translates a numpy array from decoded ("image space", float: mostly 0.0..1.0)
        to encoded ("stream space", uint8: 0..255) representation.
        '''
        msg(f'applying Encode: {Decode}')
        INTERPOLATE = lambda x, xmin, xmax, ymin, ymax: ymin + ((x - xmin) * (ymax - ymin)) / (xmax - xmin)
        if array.ndim == 2: array = np.dstack([array])
        N = array.shape[-1] # number of colors
        assert len(Decode) == 2*N
        for i in range(N):
            array[...,i] = INTERPOLATE(array[...,i].astype(float), Decode[2*i], Decode[2*i + 1], 0, 255)
        array = np.clip(np.round(array), 0, 255).astype(np.uint8)
        return array if N > 1 else array[...,0]


# ========================================================================== class PdfDecodedImage

class PdfImage(AttrDict):

    # To do:
    # 1) Lab color
    # 2) color reductions

    formats = ['array', 'pil', 'tiff', 'jbig2', 'jpeg', 'jp2', 'png']

    def __init__(self,
                 obj:IndirectPdfDict = None,
                 pil:Image.Image = None,
                 array:np.ndarray = None,
                 jp2:bytes = None):
        '''
        Creates a PdfImage.
        '''
        if sum(1 for x in [obj, pil, array,jp2] if x is not None) != 1:
            raise ValueError(f'exactly one of the [pil, array, obj] arguments should be given to constructor')
        
        self.pil = pil
        self.array = array
        self.jp2 = None
        self.dpi = None

        self.Decode = None
        self.Mask = None
        self.SMask = None

        self.isRendered = False

        if obj:

            assert obj.Subtype in ['/Image', PdfName.Image]
            if obj.stream == None: raise ValueError(f'image has no stream: {obj}')

            obj = PdfFilter.uncompress(obj)

            self.Decode = obj.Decode
            self.ColorSpace = PdfImage.xobject_get_cs(obj)
            self.bpc = PdfImage.xobject_get_bpc(obj)

            def toBytes(s:str): return s.encode('Latin-1')
            stream = toBytes(obj.stream)

            if obj.Filter == None:

                width, height = int(obj.Width), int(obj.Height)
                cpp = self.get_cpp()

                bpc_implied = (len(stream) * 8) / (width * height * cpp)
                if bpc_implied != self.bpc and int(bpc_implied) == bpc_implied:
                    warn(f'replacing bad image xobject\'s bpc = {self.bpc} with the implied bpc = {int(bpc_implied)}')
                    self.bpc = int(bpc_implied)

                array = PdfFilter.unpack_pixels(stream, width, cpp, self.bpc, truncate = True)
                if array.shape[0] > height:
                    warn(f'more rows in stream than expected: {array.shape[0]} vs {height}; truncating')
                    array = array[:height]
                if array.shape[0] != height:
                    raise ValueError(f'image height mismatch: expected {height}, read {array.shape[0]}, bpc = {self.bpc}')
                array = SImage.normalize(array)
                if self.bpc == 1: array = array.astype(bool)
                self.set_array(array)

            elif obj.Filter == '/CCITTFaxDecode':

                header = PdfImage._tiff_make_header(obj)
                self.set_pil(Image.open(BytesIO(header + stream)))

            elif obj.Filter == '/JBIG2Decode':

                jbig2_locals = stream
                try: jbig2_globals = toBytes(PdfFilter.uncompress(obj.DecodeParms.JBIG2Globals).stream)
                except: jbig2_globals = None
                self.set_pil(PdfImage._jbig2_read([jbig2_locals, jbig2_globals]))

            elif obj.Filter == '/DCTDecode': 

                pil = Image.open(BytesIO(stream))

                self.set_pil(pil)

            elif obj.Filter == '/JPXDecode':

                self.read_jp2_to_array(stream)

            else:
                raise ValueError(f'bad filter: {obj.Filter}')

            if obj.DecodeParms and obj.Filter in ['/DCTDecode', '/JPXDecode', None]:
                warn(f'ignoring /DecodeParms: {obj.DecodeParms}')

            # Handle transparency
            
            self.Mask  = obj.Mask
            self.SMask = obj.SMask

            # Invert /DeviceCMYK & /DeviceN JPEGs
            # cs_name = PdfColorSpace.get_name(self.ColorSpace)
            if self.pil and self.pil.mode == 'CMYK' and self.pil.format == 'JPEG':
                msg(f'inverting CMYK JPEG after decoding')
                self.set_pil(ImageChops.invert(self.pil))

        else:

            if pil:

                # Handle transparency
                if pil.mode in ['PA','LA','RGBA']:
                    self.pil, alpha = PdfImage._pil_split_alpha(pil)
                    self.SMask = PdfImage(pil = alpha).encode()

                self.ColorSpace = PdfImage._pil_get_colorspace(self.pil)
                self.bpc = 1 if self.pil.mode == '1' else 8
                self.dpi = self.pil.info.get('dpi')

            if jp2:

                self.read_jp2_to_array(jp2)
                self.dpi = Image.open(BytesIO(jp2)).info.get('dpi')

            if array is not None:
                SImage.validate(array)
                N = SImage.nColors(array)
                self.ColorSpace = PdfName.DeviceGray if N == 1 \
                                    else PdfName.DeviceRGB if N == 3 \
                                    else PdfName.DeviceCMYK if N == 4 \
                                    else None
                self.bpc = 1 if SImage.isBitonal(array) else 8
 
    # -------------------------------------------------------------------------------------- Basic functions

    def w(self):
        '''Returns width'''
        return self.array.shape[1] if self.array is not None else self.pil.width

    def h(self):
        '''Returns height.'''
        return self.array.shape[0] if self.array is not None else self.pil.height

    def get_cpp(self):
        '''Returns components per pixel.'''
        return PdfColorSpace.get_cpp(self.ColorSpace)

    def get_mode(self):
        '''Returns a PIL mode that is appropriate for the current colorspace and bpc values.'''
        return PdfColorSpace.get_mode(self.ColorSpace, self.bpc)

    def to_bytes(self):
        '''Returns pixel intensities as a bytes object.'''
        return self.pil.tobytes() if self.pil else PdfFilter.pack_pixels(self.array, self.bpc)

    def invert(self):
        '''Inverts image'''
        if self.pil: self.set_pil(ImageChops.invert(self.pil))
        elif self.array is not None: self.set_array(SImage.invert(self.array))
        else: raise ValueError(f'bad PdfImage: {self}')

    def __str__(self):
        '''Returns a string representation of the image that contains
        various image parameters in a compact format'''
        attrs = [a for a in ['pil','array','jp2','mask'] if self.__getattr__(a) is not None]
        if self.isRendered: attrs.append('isRendered')
        attrs = '/'.join(attrs)
        return f'{self.w()}x{self.h()}, DPI={self.dpi}, CS={PdfColorSpace.get_name(self.ColorSpace)}' \
                 + f', BPC={self.bpc}, STATE: {attrs}'


    # -------------------------------------------------------------------------------------- render()

    def render(self, pdfPage:PdfDict = None):
        '''
        '''
        if self.isRendered:
            warn('image already rendered; skipping')
            return

        msg('rendering started')

        msg(f'applying image colorspace: {PdfColorSpace.toStr(self.ColorSpace)}')
        self.apply_colorspace(self.ColorSpace, self.SMask) # Mask is needed to unmultiply alpha if necessary

        # Apply page default colorspace
        try:
            # This will fail if cs is not among the entries of the page default colorspaces dict
            cs2cs = {'/DeviceGray':'/DefaultGray', '/DeviceRGB':'/DefaultRGB', '/DeviceCMYK':'/DefaultCMYK'}
            default_cs = pdfPage.Resources.ColorSpace[cs2cs[self.ColorSpace]]
            msg(f'applying page default colorspace: {PdfColorSpace.toStr(default_cs)}')
            self.apply_colorspace(default_cs)
        except: pass

        self.isRendered = True
        msg('rendering ended')

    # -------------------------------------------------------------------------------------- apply_colorspace()

    def apply_colorspace(self, cs:CS_TYPE, mask:PdfDict = None):
        '''
        '''
        actualDecode  = PdfDecodeArray.get_actual(self.Decode, cs, self.bpc)
        defaultDecode = PdfDecodeArray.get_default(cs, self.bpc)

        cs_name = PdfColorSpace.get_name(cs)

        if cs_name == '/Indexed':

            # In the /Indexed color space, all color space transformations are performed on the palette.
            # Note: palette_cs is never /Indexed (see Adobe PDF Ref.)

            # PDF standard allows for remapping of palette indices via Decode arrays (why oh why?)
            if actualDecode != defaultDecode:
                warn(f'decoding indices in the /Indexed colorspace; this is strange, check results')
                array = self.get_array()
                array = PdfDecodeArray.decode(array, actualDecode, self.bpc)
                array = np.clip(np.round(array),0,255).astype('uint8') # clip right away: no encoding step later
                self.set_array(array)
                self.bpc = 8
                self.Decode = None
                
            # Transform the palette
            palette_cs, palette_array = PdfColorSpace.get_palette(cs)

            if PdfColorSpace.get_name(palette_cs) in ['/Separation', '/DeviceN', '/NChannel']:

                # Apply palette_cs to palette_array; both palette_array & palette_cs can change
                palette_decode = PdfDecodeArray.get_default(palette_cs, 8)
                palette_array = np.array([palette_array])
                palette_cs, palette_array = PdfColorSpace.reduce(palette_cs, palette_array, palette_decode, 8, mask)
                palette_array = palette_array[0]

                # Create new /Indexed colorspace
                self.ColorSpace = PdfArray([PdfName.Indexed, palette_cs, palette_array.shape[0], palette_array.tobytes()])

            # Reduce palette; comment this of if you do not want render() to reduce palettes
            self.ColorSpace = palette_cs
            self.set_array(palette_array[self.get_array()])
            self.bpc = 8
            self.Decode = None

        elif cs_name in ['/Separation', '/DeviceN', '/NChannel'] \
                or actualDecode != defaultDecode \
                or mask != None and mask.Matte != None:

            self.ColorSpace, array = PdfColorSpace.reduce(cs, self.get_array(), actualDecode, self.bpc, mask)
            self.set_array(array)
            self.bpc = 8
            self.Decode = None

    # -------------------------------------------------------------------------------------- change_mode()

    def change_mode(self, mode:str, intent:str = '/Perceptual'):
        '''
        Changes image mode. The mode argument can be any one of: 'L', 'RGB', 'CMYK'
        '''
        assert self.isRendered

        if mode not in ['L', 'RGB', 'CMYK']: raise ValueError(f'bad mode: {mode}')

        if mode == self.get_mode() and self.ColorSpace in [PdfName.DeviceGray, PdfName.DeviceRGB, PdfName.DeviceCMYK]:
            warn(f'old and new modes are the same: {mode}')
            return False
        
        if self.get_mode() == '1':
            if mode == 'L':
                if self.array: self.set_array(SImage.toGray(self.array))
                elif self.pil: self.set_pil(self.pil.convert('L'))
                else: raise ValueError(f'bad PdfImage: {self}')
                return True
            elif mode == 'RGB':
                if self.array: self.set_array(SImage.toColor(self.array))
                elif self.pil: self.set_pil(self.pil.convert('RGB'))
                else: raise ValueError(f'bad PdfImage: {self}')
                return True
            else:
                warn(f'cannot convert {"1"} --> {mode}')
                return False


        # Render
        pil = self.get_pil()

        # Get intent
        intents = {'/Perceptual': ImageCms.Intent.PERCEPTUAL,
                   '/RelativeColorimetric': ImageCms.Intent.RELATIVE_COLORIMETRIC,
                   '/Saturation': ImageCms.Intent.SATURATION,
                   '/AbsoluteColorimetric': ImageCms.Intent.ABSOLUTE_COLORIMETRIC }
        renderingIntent = intents.get(intent, ImageCms.Intent.PERCEPTUAL)

        msg(f'converting {pil.mode} --> {mode} (intent = {intent})')

        # Standard sRGB & CMYK profiles
        cmyk_profile = ImageCms.ImageCmsProfile(BytesIO(CMYK_DEFAULT_ICC_PROFILE))
        srgb_profile = ImageCms.createProfile('sRGB')

        # Input profile
        icc_profile = PdfColorSpace.get_icc_profile(self.ColorSpace)
        input_profile = ImageCms.ImageCmsProfile(BytesIO(icc_profile)) if icc_profile \
                            else cmyk_profile if pil.mode == 'CMYK' \
                            else srgb_profile if pil.mode in ['RGB', 'P'] else None
        if not input_profile:
            warn(f'cannot convert {self.get_mode()} --> {mode}')
            return False

        # Output profile
        output_profile = cmyk_profile if mode == 'CMYK' else srgb_profile if mode == 'RGB' else None

        if not output_profile:
            # Conversion to L: first convert to RGB to account for input_profile, then to L
            # image = ImageCms.profileToProfile(image, input_profile, srgb_profile,
            #                                 renderingIntent = renderingIntent, outputMode='RGB')
            pil = pil.convert('L')
        else:
            pil = ImageCms.profileToProfile(pil, input_profile, output_profile,
                                            renderingIntent = renderingIntent, outputMode=mode)
        
        if 'icc_profile' in pil.info: del pil.info['icc_profile']

        self.ColorSpace = PdfName.DeviceGray if pil.mode == 'L' \
                            else PdfName.DeviceRGB if pil.mode == 'RGB' \
                            else PdfName.DeviceCMYK
                            # else PdfColorSpace.make_icc_based_colorspace(CMYK_DEFAULT_ICC_PROFILE, N=4)

        self.set_pil(pil)

    # -------------------------------------------------------------------------------------- get_pil()

    def get_pil(self):
        '''
        Creates a PIL Image representation in self.pil, if not yet created, and returns it.
        '''
        if not self.pil:
            assert self.array is not None and self.ColorSpace != None and self.bpc != None
            msg('array -> pil')
            mode = PdfColorSpace.get_mode(self.ColorSpace, self.bpc)
            # A PIL bug: https://stackoverflow.com/questions/50134468/convert-boolean-numpy-array-to-pillow-image
            self.pil = Image.fromarray(self.array, mode = mode) if mode != '1' \
                else Image.frombytes(mode='1', size=self.array.shape[::-1], data=np.packbits(self.array, axis=1))
        return self.pil

    # -------------------------------------------------------------------------------------- get_array()

    def get_array(self):
        '''
        Creates a numpy array representation in self.array, if not yet created, and returns it.
        '''
        if self.array is None:
            assert self.pil
            msg('pil -> array')
            self.array = np.array(self.pil)
        SImage.validate(self.array)
        return self.array

    # -------------------------------------------------------------------------------------- set_pil()

    def set_pil(self, pil:Image.Image):
        '''
        Sets self.pil to the given PIL image, and self.array & self.jp2 to None
        '''
        self.pil = pil
        self.array = None
        self.jp2 = None

    # -------------------------------------------------------------------------------------- set_array()

    def set_array(self, array:np.ndarray):
        '''
        Sets self.array to the given numpy image array, and self.pil & self.jp2 to None
        '''
        self.pil = None
        self.array = array
        self.jp2 = None

    # -------------------------------------------------------------------------------------- set_jp2()

    def read_jp2_to_array(self, jp2:bytes):
        '''
        Reads jp2 into self.array and sets self.bpc.
        If self.ColorSpace is None it is set to the internal colorspace of the jp2 image.
        Also, self.jp2 is set to jp2, unless self.ColorSpace == ['/Indexed', ..].

        Having self.array, self.ColorSpace and self.bpc set allows for subsequent image analysis/modification.
        This leaves self.jp2 as a kind of storage for the original JPEG2000 file's bytes
        in case we want to save the jp2 image data without re-encoding.

        Note: when self.ColorSpace is an /Indexed colorspace, self.array will be read as
        a /DeviceRGB or /DeviceCMYK colorspace, and therefore, and therefore 1) self.ColorSpace
        will be set to /DeviceRGB or /DeviceCMYK to reflect that; 2) jp2 will not be stored
        in self.jp2.
        '''
        self.array, jp2_cs, self.bpc = PdfImage._jp2_read(jp2)
        cs = self.ColorSpace
        isIndexed = isinstance(cs, PdfArray) and cs[0] == '/Indexed'
        if not isIndexed:
            self.jp2 = jp2
        if not cs or isIndexed:
            self.ColorSpace = jp2_cs # obj.ColorSpace overrides internal JPX colorspace
        self.pil = None

    # -------------------------------------------------------------------------------------- resize()

    def resize(self, width:int, height:int):
        '''
        Resize the image to the given dimensions.
        '''
        self.set_pil(self.get_pil().resize((width, height)))

    # -------------------------------------------------------------------------------------- saveAs

    def saveAs(self, Format:str,
                    Q:float = None,
                    CR:float = None,
                    optimize = True,
                    render = False,
                    intent:str = '/Perceptual',
                    embedProfile:bool = True,
                    invertCMYK = False):
        '''
        Returns a tuple `(imageStream[bytes], fileExtension[str])`.
        The encoding is done in the given `Format` using the
        `Q` (quality), `CR` (compression ratio) and `optimize` parameters.

        The image has to be rendered first in order to produce a faithful representation.
        '''
        ext = {'TIFF':'.tif', 'PNG':'.png', 'JPEG':'.jpg', 'JPEG2000':'.jp2', 'JBIG2':'jb2'}

        addAlpha = render and (self.Mask or self.SMask)

        if Format == 'JPEG' and addAlpha:
            raise ValueError(f'cannot save an image transparency as JPEG')

        if Format == 'auto':
            Format = self.pil.format if self.pil and self.pil.format in ext \
                        else 'JPEG2000' if self.jp2 \
                        else 'TIFF'
            if addAlpha and Format == 'JPEG':
                Format = 'TIFF'

        if Format not in ext: raise ValueError(f'cannot save in format: {Format}')

        if addAlpha:
            msg('rendering alpha')
            alpha = PdfImage(obj = self.Mask or self.SMask)
            alpha.render()
            if alpha.w() != self.w() or alpha.h() != self.h():
                if self.w() * self.h() > alpha.w() * alpha.h():
                    msg(f'resizing alpha to fit image')
                    alpha.resize(self.w(), self.h())
                else:
                    msg(f'resizing image to fit alpha')
                    self.resize(alpha.w(), alpha.h())

            # # ???????????????????????????????
            # if alpha.mode == '1':
            #     alpha = alpha.convert('L')

        if Format == 'JPEG2000':

            if self.jp2:
                stream  = self.jp2
            else:
                assert Q or CR
                if Q: assert 0 < Q <= 100
                if CR: assert CR > 1
                stream = PdfImage._jp2_write(array = self.get_array(),
                                            alpha = alpha.get_array() if addAlpha else None,
                                            cs = self.ColorSpace,
                                            Q = Q, CR = CR)
        else:
 
            # Create pil
            pil = self.get_pil()

            if render:
                msg('rendering pil')
                self.render()
                if addAlpha and self.get_mode() not in ['L', 'RGB']:
                    self.change_mode('RGB' if self.get_mode() != '1' else 'L', intent)
                pil = self.get_pil()
                PdfImage._pil_set_colorspace(pil, self.ColorSpace)

                if addAlpha:
                    msg('adding alpha')
                    # alpha.render()
                    assert alpha.get_cpp() == 1
                    alpha.change_mode('RGB' if alpha.get_mode() != '1' else 'L', intent)
                    alpha_pil = alpha.get_pil()
                    # invert bitonal masks
                    if self.Mask and self.Mask.ImageMask == PdfObject('true'):
                        alpha_pil = ImageChops.invert(alpha_pil)
                    PdfImage._pil_set_colorspace(alpha_pil, alpha.ColorSpace)
                    pil.putalpha(alpha_pil)

            if invertCMYK and pil.mode == 'CMYK' and Format == 'JPEG':
                msg('inverting CMYK JPEG before encoding')
                pil = ImageChops.invert(pil)

            bs = BytesIO()
            Q = Q if Q else 'keep' if pil.format == 'JPEG' and Format == 'JPEG' else 100
            msg(f'saving with Format = {Format}, Q = {Q}')
            # tiff_compression = 'group4' if pil.mode == '1' else 'tiff_lzw' if Format == 'TIFF' else None
            tiff_compression = 'group4' if pil.mode == '1' else 'tiff_adobe_deflate' if Format == 'TIFF' else None
            icc_profile = pil.info.get('icc_profile') if pil.info and embedProfile else None

            dpi = self.dpi if self.dpi not in (None, (1,1)) else (72,72)

            if Format == 'JPEG':
                pil.save(bs, Format, quality=Q, optimize = optimize, icc_profile = icc_profile, info = pil.info, dpi=dpi)
            elif Format == 'PNG':
                pil.save(bs, Format, optimize = optimize, icc_profile = icc_profile)
            elif Format == 'TIFF':
                pil.save(bs, Format, compression = tiff_compression, icc_profile = icc_profile)
            else:
                raise ValueError(f'format not supported in saveAs(): {Format}')
 
            stream = bs.getvalue()

            # A bug found in some JPEG files
            if Format == 'JPEG' and stream[-2:] != b'\xff\xd9':
                raise ValueError('bad JPEG stream')
                # stream += b'\xff\xd9'

        return stream, ext[Format]


    # -------------------------------------------------------------------------------------- to_pdf_page()

    def to_pdf_page(self):
        '''
        Convert the image to a PDF page.
        '''
        # Encode image
        xobj = self.encode()
        print('RESULT:')
        pprint(xobj)

        dpi = self.dpi if self.dpi not in [None, (1,1)] else (72,72)

        # Create page
        w,h = self.w(), self.h()
        w,h = w * 72/dpi[0], h*72/dpi[1]
        p = lambda x: round(x*1000000)/1000000
        q = lambda x: f'{p(x):f}'.rstrip('0').rstrip('.')
        page = IndirectPdfDict(
            Type = PdfName.Page,
            MediaBox = [0, 0, p(w), p(h)],
            Contents = IndirectPdfDict(stream=f'{q(w)} 0 0 {q(h)} 0 0 cm\n/Im1 Do\n'),
            Resources = PdfDict(XObject = PdfDict(Im1 = xobj))
        )

        return page

    # -------------------------------------------------------------------------------------- compress()

    @staticmethod
    def recompress(image_obj:IndirectPdfDict, Format:str = None, Q:float = None, CR:float = None):
        '''
        (Re-)compress image stream. The format can be any one of:
        
        * 'ZIP': compress with /FlateDecode without PNG predictors
        * 'PNG': compress with /FlateDecode with PNG predictors
        * 'JPEG': compress with /DCTDecode; compressionRatio should be specified
        * 'JPEG2000': compress with /JPXDecode; either compressedQuality or compressionRatio should be specified

        The obj.stream is replaced only if the (re-)compression decreases stream size. An exception to this
        is the 'ZIP' compression format, for which the stream is always (re-)compressed, even if
        its size increases as a result. Therefore use 'ZIP' to (re-)compress only if you need
        the /FlateDecode filter for compatibility/tests. For stream size reduction with
        /FlateDecode try 'PNG' compression instead.

        Returns True/False depending on whether image stream has been replaced.
        '''

        def print_size_change(size_old, size_new):
            return f'Size: {size_old} --> {size_new} ({((size_new - size_old)*100)//size_old:+d}%)'

        size_old = int(image_obj.Length)

        obj = PdfFilter.uncompress(image_obj)
        image = PdfImage(obj = obj)

        if PdfColorSpace.get_name(image.ColorSpace) == '/Indexed':
            msg('/Indexed colorspace will not be recompressed')
            return False
        
        bpc = image.bpc
        cs = image.ColorSpace

        DecodeParms = None
 
        if Format == 'ZIP':

            stream = zlib.compress(image.to_bytes())
            Filter = PdfName('FlateDecode')

        elif Format == 'PNG':

            if image.get_mode() not in ['L', 'RGB']:
                msg(f'mode {image.get_mode()} not supported by PNG; skipping')
                return False

            png, _ = image.saveAs('PNG')
            stream, Predictor = PdfImage._png_get_stream_and_filter(png)
            Filter = PdfName('FlateDecode')
            DecodeParms = PdfDict(Predictor = Predictor, # PDF Spec.: actual value doesn't matter a.l.a. it's 10..15
                                    Colors = PdfColorSpace.get_cpp(cs),
                                    BitsPerComponent = bpc,
                                    Columns = image.w())

        elif Format == 'JPEG':

            stream, _ = image.saveAs('JPEG', Q, invertCMYK = True)
            Filter = PdfName('DCTDecode')

        elif Format == 'JPEG2000':

            assert Q or CR
            if Q: assert 0 < Q <= 100
            if CR: assert CR > 1

            stream, _ = image.saveAs('JPEG2000', Q, CR)
            Filter = PdfName('JPXDecode')

        else:
            raise ValueError(f'bad Format: {Format}')

        # Compare sizes before/after
        size_new = len(stream)
        if Format == 'ZIP' or size_new * 10 < size_old * 9:
            msg(f're-compressing {obj.Filter} --> {Filter}: {print_size_change(size_old, size_new)}')
            obj.stream = stream.decode('Latin-1')
            obj.Filter = Filter
            obj.DecodeParms = DecodeParms
            obj.BitsPerComponent = bpc
            obj.ColorSpace = cs

            PdfImage.xobject_copy(obj, image_obj)

            return True
        else:
            msg(f'skipping {obj.Filter} --> {Filter}: {print_size_change(size_old, size_new)}')
            return False

    # -------------------------------------------------------------------------------------- encode()

    def encode(self):
        '''
        '''
        DecodeParms = None
        Decode = None

        if self.bpc == 1:

            Filter = 'CCITTFaxDecode'
            pil = self.get_pil()
            stream = PdfImage._tiff_get_strip(PdfImage._tiff_write(pil))
            DecodeParms = PdfDict(K = -1,
                                  Columns = self.w(),
                                  Rows = self.h(),
                                  BlackIs1 = PdfObject('true'))
            
        elif self.jp2:

            Filter = 'JPXDecode'
            stream = py23_diffs.convert_load(self.jp2)

        elif not self.pil or self.pil.format in ['RAW', 'GIF', 'PNG', 'TIFF', 'JPEG2000', None]:

            Filter = 'FlateDecode'
            stream = zlib.compress(self.to_bytes())
 
        elif self.pil.format == 'JPEG':

            Filter = 'DCTDecode'
            stream, _ = self.saveAs('JPEG', embedProfile = False, invertCMYK = True)

        else:
            raise ValueError(f'unsupported format: {self.pil.format}')

        return IndirectPdfDict(
            Type = PdfName.XObject,
            Subtype = PdfName.Image,
            Width = self.w(),
            Height = self.h(),
            BitsPerComponent = self.bpc,
            ColorSpace = self.ColorSpace,
            Filter = PdfName(Filter),
            DecodeParms = DecodeParms,
            Decode = Decode or self.Decode,
            Mask = self.Mask,
            SMask = self.SMask,
            stream = py23_diffs.convert_load(stream)
        )


    # -------------------------------------------------------------------------------------- jbig2_encode()

    # @staticmethod
    # def jbig2_encode():
    #     '''
    #     '''

    #     # stream = self.jbig2[0]
    #     # Filter = 'JBIG2Decode'
    #     # if self.jbig2[1]:
    #     #     globals = IndirectPdfDict(Filter = PdfName.FlateDecode, stream = zlib.compress(self.jbig2[1]))
    #     # DecodeParms = PdfDict(JBIG2Globals = globals)

    # -------------------------------------------------------------------------------------- xobject_from_inline_image()

    @staticmethod
    def xobject_from_inline_image(inline_img:list):
        '''Converts a PDF inline image (represented as an element of a parsed PDF stream tree: ['BI', [{header_dict}, "data_str"]])
        to a PDF image object. This helps treat both inline and object images in PDF in a uniform way.
        '''
        if inline_img[0] != 'BI': return None

        header, data = inline_img[1]

        data = data.strip(' \n\r')
        if data[:2] != 'ID' or data[-2:] != 'EI': return None
        data = data[3:-2]
        data = data.rstrip('\r\n')
        # data = data.rstrip('\n') if data[-1] == '\n' else data.rstrip('\r')

        filter_name = get_any_key(header, '/F','/Filter')
        filter_name_aliases = {'/AHx':'/ASCIIHexDecode', '/A85':'/ASCII85Decode', '/LZW':'/LZWDecode', '/Fl':'/FlateDecode',
                                '/RL':'/RunLengthDecode', '/CCF':'/CCITTFaxDecode', '/DCT':'/DCTDecode'}
        if filter_name in filter_name_aliases: filter_name = filter_name_aliases[filter_name]

        obj = IndirectPdfDict(
            Type = '/XObject',
            Subtype = '/Image',
            Width = get_any_key(header, '/W','/Width'),
            Height = get_any_key(header, '/H','/Height'),
            Filter = filter_name,
            DecodeParms = get_any_key(header, '/DP', '/DecodeParms'),
            BitsPerComponent = get_any_key(header, '/BPC','/BitsPerComponent'),
            ColorSpace = get_any_key(header, '/CS','/ColorSpace'),
            Decode = get_any_key(header, '/D','/Decode'),
            ImageMask = PdfObject(get_any_key(header, '/IM','/ImageMask')),
            Intent = get_any_key(header, '/Intent'),
            Interpolate = get_any_key(header, '/I','/Interpolate')
        )
        obj.stream = data

        return obj

    # -------------------------------------------------------------------------------------- xobject_to_inline_image_stream()

    @staticmethod
    def xobject_to_inline_image_stream(obj:PdfObject):
        '''
        Converts a PDF image xobject to a PDF inline image stream.
        '''
        p = lambda z: '<<'+' '.join(f'{x} {y}' for (x,y) in z.items())+'>>' if isinstance(z,PdfDict) \
            else '['+' '.join(f'{x}'for x in z)+']' if isinstance(z,PdfArray) \
            else f'{z}'
         
        if obj == None or obj.Subtype != PdfName.Image: return None

        s = f'BI\n/W {obj.Width} /H {obj.Height}'
        if obj.Filter != None: s += f' /F {obj.Filter}'
        if obj.DecodeParms != None: s += f' /DP ' + p(obj.DecodeParms)
        if obj.BitsPerComponent != None: s += f' /BPC {obj.BitsPerComponent}'
        if obj.ColorSpace != None: s += f' /CS {obj.ColorSpace}'
        if obj.Decode != None: s += f' /D {obj.Decode}'
        if obj.ImageMask != None: s += f' /IM {obj.ImageMask}'
        if obj.Intent != None: s += f' /Intent {obj.Intent}'
        if obj.Interpolate != None: s += f' /I {obj.Interpolate}'
        s += '\nID\n'
        s += obj.stream
        s += '\nEI\n'

        return s

    # -------------------------------------------------------------------------------------- xobject_get_bpc()

    @staticmethod
    def xobject_get_bpc(obj:PdfDict):
        '''
        Returns bits per pixel
        '''
        if obj.Subtype != PdfName.Image: raise ValueError(f"Not an image: {obj}")
        return 1 if obj.ImageMask == PdfObject('true') \
            else None if obj.Filter == PdfName.JPXDecode \
            else int(obj.BitsPerComponent)

    # -------------------------------------------------------------------------------------- xobject_get_colorspace()

    @staticmethod
    def xobject_get_cs(obj:PdfDict) -> CS_TYPE:
        '''
        Returns obj.ColorSpace if it's set, or /DeviceGray if obj.ImageMask == true,
        or None otherwise.
        '''
        if obj.Subtype != PdfName.Image: raise ValueError(f"Not an image: {obj}")
        return obj.ColorSpace if obj.ColorSpace != None \
            else PdfName.DeviceGray if obj.ImageMask == PdfObject('true') \
            else None

    # -------------------------------------------------------------------------------------- xobject_copy()

    @staticmethod
    def xobject_copy(obj_from:IndirectPdfDict, obj_to:IndirectPdfDict):
        '''
        '''
        # obj_to.clear()
        for k,v in obj_from.items():
            obj_to[k] = v
        obj_to.stream = obj_from.stream


    # -------------------------------------------------------------------------------------- xobject_copy()

    @staticmethod
    def xobject_make_transparency_group_graphics_state(obj:IndirectPdfDict):
        '''
        Make an extended graphics state with the image obj used as a source of the transparency
        alpha. Use it like this:
        ```
        GS1 = PdfImage.xobject_make_transparency_group_graphics_state(obj)
        xobject.Resources.ExtGState = PdfDict(GS1 = GS1)
        xobject.stream += 'q /GS1 gs [graphics operators] Q'
        ```
        '''
        return  PdfDict(
                    Type = PdfName.ExtGState,
                    BM = PdfName.Multiply,
                    CA = 1, ca = 1,
                    SMask = PdfDict(
                        BC = [0,0,0], # Backdrop color (black)
                        S = PdfName.Luminosity, # source of mask's alpha
                        # S = PdfName.Alpha, # source of mask's alpha

                        # Transparency group XObject to be used as mask
                        G = IndirectPdfDict(
                            Type = PdfName.XObject,
                            Subtype = PdfName.Form,
                            BBox = [0,0,1,1],
                            Group = IndirectPdfDict(
                                CS = PdfName.DeviceRGB,
                                I = PdfObject('true'), # Isolated, i.e. mask's background is transparent
                                K = PdfObject('false'), # Knockout is false, see PDF Ref
                                S = PdfName.Transparency
                            ),
                            Resources = PdfDict(XObject = PdfDict(mask = obj)),
                            stream = '/mask Do\n'
                        )
                    )
                )

    # -------------------------------------------------------------------------------------- _jbig2_read()

    @staticmethod
    def _jbig2_read(jbig2:tuple[bytes, bytes]):
        '''
        Returns a PIL image
        '''
        loc, glob = jbig2
        with tempfile.TemporaryDirectory() as tmp:
            T = lambda fileName: os.path.join(tmp, fileName)
 
            open(T('locals'),'wb').write(loc) 
            if glob:
                open(T('globals'),'wb').write(glob)
                OS.execute(['jbig2dec', '-e', '-o', T('out.png'), T('globals'), T('locals')])
            else:
                OS.execute(['jbig2dec', '-e', '-o', T('out.png'), T('locals')])
            
            return Image.open(T('out.png'))

        # If you have to re-assemble the JBIG2 file for some reason, just add the header:
        # 1-page, sequential order; see JBIG2 Specs
        # header = b'\x97\x4A\x42\x32\x0D\x0A\x1A\x0A\x01\x00\x00\x00\x01'

    # -------------------------------------------------------------------------------------- _jp2_read()

    @staticmethod
    def _jp2_read(jp2:bytes):
        '''
        Reads a JPEG2000 file (a bytes object) and returns a tuple (numpyArray, colorSpace, bitsPerComponent).
        '''
        with tempfile.TemporaryDirectory() as tmp:
            T = lambda fileName: os.path.join(tmp, fileName)

            open(T('temp.jp2'),'wb').write(jp2)
            jp2k = Jp2k(T('temp.jp2'))          # call print(jp2k) to inspect the boxes

            # Get the data
            # print(jp2k)
            try: bpc = int(jp2k.box[2].box[0].bits_per_component)
            except: bpc = int(jp2k.box[3].box[0].bits_per_component)

            data = jp2k[:] # This parses all JPEG2000 image slices for a full resolution image
            if bpc == 1: data = data.astype(bool)
            elif bpc <= 8: data = data.astype('uint8')
            else: raise ValueError(f'unsupported JPEG2000 bit depth: {bpc}')
            data = SImage.normalize(data)
            SImage.validate(data)

            # Get the colorspace
            try: EnumCS = jp2k.box[2].box[1].colorspace
            except: EnumCS = jp2k.box[3].box[1].colorspace

            if EnumCS != None:
                cs_name = {12:'CMYK', 16:'RGB', 17:'Gray', 18:'YCC', 20:'e-sRGB', 21:'ROMM-RGB'}.get(EnumCS)
                if cs_name in [None, 'YCC', 'e-sRGB', 'ROMM-RGB']:
                    raise ValueError(f'unsupported JPX colorspace: {cs_name}')
                cs = PdfName('Device' + cs_name)
            else:
                icc_profile = jp2k.box[3].box[1].icc_profile
                cs = PdfColorSpace.make_icc_based_colorspace(icc_profile, N = SImage.nColors(data))

        return data, cs, bpc

    # -------------------------------------------------------------------------------------- _jp2_write()

    @staticmethod
    def _jp2_write(array:np.ndarray, alpha:np.ndarray, cs:CS_TYPE, Q:int, CR:float):
        '''
        Encode array as a JPEG2000 image.
        '''
        assert CR or Q
        CR = int(round(CR)) if CR != None else int(round(2 ** ((100 - Q)/10.0)))
        cpp = PdfColorSpace.get_cpp(cs)
        if cpp not in [1,3]:
            raise ValueError(f'Jp2k: cpp = {cpp} not supported (must be 1 or 3)')
        with tempfile.TemporaryDirectory() as tmp:
            T = lambda fileName: os.path.join(tmp, fileName)
            # can also try: cratios=[CR*4, CR*2, CR]
            Jp2k(T('encoded.jp2'), array, colorspace='RGB' if cpp == 3 else 'Gray', cratios=[CR])
            
            return open(T('encoded.jp2'), 'rb').read()

    # -------------------------------------------------------------------------------------- _tiff_make_header()

    @staticmethod
    def _tiff_make_header(obj:PdfDict):
        '''
        Make a TIFF header.

        The CCITTFaxDecode -encoded stream is essentially a TIFF file without the header, so
        this function reconstructs the missing TIFF file header.
        '''
        assert obj.Subtype in ['/Image', PdfName.Image] and obj.Filter == '/CCITTFaxDecode'
 
        # ------------------------------------------------------------------ Get info

        parm = obj.DecodeParms
        width, height = int(obj.Width), int(obj.Height)
        length = int(obj.Length)

        K = get_key(parm, '/K', '0')
        tiff_compression = 3 if int(K) >= 0 else 4

        BlackIs1 = get_key(parm, '/BlackIs1', 'false')
        # tiff_photometric = 0 if BlackIs1 == 'true' else 1
        tiff_photometric = 1 if BlackIs1 == 'true' else 0           # ??? Why is this backwards ???
        # if obj.ImageMask == PdfObject('true'): tiff_photometric = 1 - tiff_photometric

        bpc = 1
        cpp = 1

        predictor = 1

        encodedByteAlign = get_key(parm, '/EncodedByteAlign') == 'true'
        if encodedByteAlign:
            warn(f'*** /EncodedByteAlign == true, this is in beta-testing, check results ***')
            if tiff_compression != 3:
                raise ValueError(f'/EncodedByteAlign == true is not supported for /CCITTFaxDecode Group4 (T6) compression')

        tiff_header_struct = '<' + '2s' + 'H' + 'L' + 'H' + 'HHLL' * 11 + 'L'
        tiff_header = struct.pack(tiff_header_struct,
                        b'II',  # Byte order indication: Little indian
                        42,  # Version number (always 42)
                        8,  # Offset to first IFD
                        8,  # --- IFD starts here; the number of tags in IFD
                        256, 4, 1, width,  # ImageWidth, LONG, 1, width
                        257, 4, 1, height,  # ImageLength, LONG, 1, length
                        258, 3, 1, bpc,  # BitsPerSample, SHORT, 1, 1; this is the default, so omit?
                        259, 3, 1, tiff_compression,  # Compression, SHORT, 1, 3 = Group 3, 4 = Group 4, 8 - Deflate (ZIP)
                        262, 3, 1, tiff_photometric,  # Photometric interpretation, SHORT, 1, 0 = WhiteIsZero
                        273, 4, 1, struct.calcsize(tiff_header_struct),  # StripOffsets, LONG, 1, len of header
                        277, 3, 1, cpp,  # SamplesPerPixel, LONG, 1, component per pixel (cpp)
                        278, 4, 1, height,  # RowsPerStrip, LONG, 1, height
                        279, 4, 1, length,  # StripByteCounts, LONG, 1, size of image
                        292, 4, 1, 4 if encodedByteAlign else 0,  # TIFFTAG_GROUP3OPTIONS, LONG, 1, 0..7 (flags)
                        317, 3, 1, predictor,  # Predictor, SHORT, 1, predictor = 1 (no prediction) or 2 (=left)
                        0  # --- IFD ends here; offset of the next IFD, 4 0-bytes b/c no next IFD
                        )
        return tiff_header

    # -------------------------------------------------------------------------------------- _tiff_write()

    @staticmethod
    def _tiff_write(pil:Image.Image):

        # Make sure Pillow produces single-strip TIFF; this works with Pillow versions < 8.3.0 or >=8.4.0  
        # See: https:/python-pillow/Pillow/pull/5744
        # For Pillow versions from 8.3.0 to 8.3.9 see:
        # https://gitlab.mister-muffin.de/josch/img2pdf/commit/6eec05c11c7e1cb2f2ea21aa502ebd5f88c5828b
        # https://gitlab.mister-muffin.de/josch/img2pdf/issues/46
        TiffImagePlugin.STRIP_SIZE = 2 ** 31

        bs = BytesIO()
        pil.save(bs, format="TIFF", compression='group4' if pil.mode == '1' else 'tiff_lzw')
        # pil.save(bs, format="TIFF", compression='group4' if pil.mode == '1' else 'tiff_adobe_deflate')

        return bs.getvalue()
    
    # -------------------------------------------------------------------------------------- _tiff_get_strip()

    @staticmethod
    def _tiff_get_strip(tiff:bytes):
        '''
        Returns the body of a single-strip TIFF file
        '''

        bs = BytesIO(tiff)
        img = Image.open(bs)

        # Read the TIFF tags to find the offset(s) and length of the compressed data strips.
        strip_offsets = img.tag_v2[TiffImagePlugin.STRIPOFFSETS]
        strip_bytes = img.tag_v2[TiffImagePlugin.STRIPBYTECOUNTS]
        rows_per_strip = img.tag_v2.get(TiffImagePlugin.ROWSPERSTRIP, 2 ** 32 - 1)
        if len(strip_offsets) != 1 or len(strip_bytes) != 1:
            raise ValueError("Expected a single strip")
        (offset,), (length,) = strip_offsets, strip_bytes

        bs.seek(offset)
        return bs.read(length)

    # -------------------------------------------------------------------------------------- _png_get_stream_and_filter()

    @staticmethod
    def _png_get_stream_and_filter(png:bytes):
        '''
        Returns a tuple (IDAT_chunk, pngFilter), where pngFilter == 10..15 is the PNG prediction filter.
        '''

        png = BytesIO(png)
        idatChunks = []
        png.seek(0)
        png.read(8)

        while True:
            length = struct.unpack(b'!L', png.read(4))[0]
            type = png.read(4)
            data = png.read(length)
            crc  = struct.unpack(b'!L', png.read(4))[0]
            if crc != zlib.crc32(type + data): warn(f'crc error in chunk: {type}'); return None
            if type == b'IDAT': idatChunks.append(data)
            if type == b'IHDR':
                assert length == 13
                width, height, bpc, color_type, compression, pngFilter, interlace = struct.unpack(b'!LLBBBBB', data)
                assert (width, height) == image.size
                assert bpc == 8
                assert color_type in [0, 2] # image.mode == 'RGB' or 'L'
                assert interlace == 0
                assert 10 <= pngFilter <= 15
            if type == b'IEND': break

        if len(idatChunks) == 0: warn(f'no IDAT chunks in PNG') ; return None
        return b''.join(chunk for chunk in idatChunks), pngFilter

    # -------------------------------------------------------------------------------------- _pil_get_colorspace()

    @staticmethod
    def _pil_get_colorspace(image:Image.Image):
        '''
        '''
        mode2cpp = {'L':1, 'RGB':3, 'CMYK':4}
        mode2cs = {'1':PdfName.DeviceGray, 'L':PdfName.DeviceGray, 'RGB':PdfName.DeviceRGB, 'CMYK':PdfName.DeviceCMYK}

        assert image.mode != 'PA' # reduce 'PA' modes before calling this function
        mode = image.mode if image.mode[-1] != 'A' else image.mode[:-1]

        cs = None

        if mode == 'P':
            # palette = image.getpalette()
            # if palette == None: return None
            # palette = b''.join(v.to_bytes(1,'big') for v in image.getpalette())

            palette_bytes = image.palette.getdata()[1]
            palette_mode = image.palette.mode
            assert palette_mode in mode2cpp

            cpp = mode2cpp.get(palette_mode)
            assert len(palette_bytes) % cpp == 0

            base = mode2cs.get(palette_mode)
            hival = len(palette_bytes) // cpp - 1
            assert hival <= 255
            palette_xobj = IndirectPdfDict(Filter = PdfName.FlateDecode,
                                            stream = zlib.compress(palette_bytes).decode('Latin-1'))
            cs = PdfArray([PdfName.Indexed, base, hival, palette_xobj])

        icc_profile = image.info.get('icc_profile')
        if icc_profile:
            if mode == 'P':
                cs[1] = PdfColorSpace.make_icc_based_colorspace(icc_profile, N = cpp)
            else:
                cpp = mode2cpp.get(mode)
                cs = PdfColorSpace.make_icc_based_colorspace(icc_profile, N = cpp)

        if not cs:
            cs = mode2cs[mode]

        return cs

    # -------------------------------------------------------------------------------------- _pil_set_colorspace()

    @staticmethod
    def _pil_set_colorspace(image:Image.Image, cs:CS_TYPE = None):
        '''
        '''
        if cs == None:
            del image.palette
            del image.info['icc_profile']
        else:
            # Palette
            palette_cs, palette_array = PdfColorSpace.get_palette(cs)
            if palette_cs != None:
                palette_mode = PdfColorSpace.get_mode(palette_cs, 8)
                assert palette_mode == 'RGB'
                assert palette_array.shape[-1] == 3
                image.putpalette(palette_array.tobytes(), rawmode='RGB')

            # ICC Profile
            if icc_profile := PdfColorSpace.get_icc_profile(cs):
                image.info['icc_profile'] = icc_profile

    # -------------------------------------------------------------------------------------- _pil_split_alpha()

    @staticmethod
    def _pil_split_alpha(image:Image.Image):
        '''
        Returns a tuple (base, alpha) if the PIL image contains an alpha channel (i.e., if its mode
        is PA, LA or RGBA), or a tuple (pil, None) otherwise.
        '''
        alpha = None

        if image.mode == 'PA':

            icc_before = image.info.get('icc_profile')
            image = image.convert('RGBA')
            assert icc_before == image.info.get('icc_profile')

        if image.mode in ['LA', 'RGBA']:

            icc_profile = image.info.get('icc_profile')
            if image.mode == 'LA':
                image, alpha = image.split()
            if image.mode == 'RGBA':
                r, g, b, alpha = image.split()
                image = Image.merge('RGB',(r,g,b))
            if 'icc_profile' in alpha.info:
                del alpha.info['icc_profile']
            if icc_profile:
                image.info['icc_profile'] = icc_profile

        return image, alpha

# ============================================================================= jbig2_compress()

def jbig2_compress(bitonal_images:dict, cpc:bool = False):
    '''
    Compress bitonal images with the JBIG2 codec
    '''
    if len(bitonal_images) == 0: warn('no images to compress') ; return
    msg('started')

    def RUN(cmdList:list):
        if not isinstance(cmdList, list): sys.exit(f'bad cmdList: {cmdList}')
        result = subprocess.run(cmdList, capture_output=True)
        if result.returncode:
            print(f'ERROR: command failed: {cmdList}')
            print(result.stdout.decode('utf-8'))
            print(result.stderr.decode('utf-8'))
            sys.exit(1)

    def intToPrefix(i:int):
        s = ''
        for _ in range(3):
            s = chr(ord('a') + i % 26) + s
            i = i // 26
        return s


    with tempfile.TemporaryDirectory() as tmp:
        T = lambda fileName: os.path.join(tmp, fileName)

        image_objects = list(bitonal_images.values())

        # Dump images to tmpDir
        tif_paths = []
        msg(f'decoding and extracting {len(bitonal_images)} bitonal images')
        for n, obj in enumerate(image_objects):
            pprint(obj)
            image = PdfImage(obj = obj)
            array = image.get_array()
            if np.mean(array) < 0.5:
                image.set_array(np.logical_not(array))
                decode = [float(x) for x in obj.Decode] if obj.Decode != None else None
                obj.Decode = None if decode == [1,0] else PdfArray(['1','0'])
                image.Decode = obj.Decode
            tif_stream, _ = image.saveAs('TIFF')
            tif_path = T(f'in-{n:04d}.tif')
            open(tif_path, 'wb').write(tif_stream)
            tif_paths.append(tif_path)

        if cpc:
            cpcCmd = ['wine', '/Users/user/Code/Win/CPCTool-530-Win32-X86_regged.exe']
            msg(f'compressing {len(tif_paths)} bitonal images with CPCTool')
            RUN(cpcCmd + tif_paths +['-o', T('cpc.cpc')])
            RUN(cpcCmd + [T('cpc.cpc'), '-o', T('cpc.tif')])
            RUN(['tiffsplit', T('cpc.tif'), T('split-')])

            for i, name in enumerate(tif_paths):
                shutil.move(T('split-') + intToPrefix(i) + '.tif', name)

        msg(f'compressing {len(tif_paths)} bitonal images with JBIG2')
        RUN(['jbig2', '-p', '-s', '-t', '0.97999', '-b', T('result')] + tif_paths)

        try:
            globals = IndirectPdfDict(stream = py23_diffs.convert_load(open(T('result.sym'),'rb').read()))
        except:
            globals = None
        for n, obj in enumerate(image_objects):
            obj.Filter = PdfName.JBIG2Decode
            obj.DecodeParms = PdfDict(JBIG2Globals = globals) if globals else None
            obj.stream = py23_diffs.convert_load(open(T(f'result.{n:04d}'),'rb').read())

    msg('ended')

# ============================================================================= modify_image_xobject()

def modify_image_xobject(image_obj:IndirectPdfDict, pdfPage:PdfDict, options:PdfDict):
    '''
    This function performs in-place modifications of an image object, such as: JPEG/JPEG2000
    compression, resizing/upsampling, color space conversions.
    '''

    # # Fix predictors
    # if options.predictors:
    #     msg('Checking images PNG predictors')
    #     _ = PdfFilter.uncompress(image_obj)
    #     if mask != None:
    #         msg('Checking mask PNG predictors')
    #         _ = PdfFilter.uncompress(mask)

    image = PdfImage(obj = image_obj)

    intent = image_obj.Intent if options.intent == 'native' else options.intent

    modified = False

    if image.bpc == 1:

        if options.scale2x:
            msg(f'upsampling with scale2x')
            image.set_array(SImage.scale2x(image.get_array())[0])
            modified = True

        if options.despeckle:
            msg(f'despeckling, threshold = {options.despeckle}')
            image.set_array(SImage.despeckle(image.get_array(), threshold = options.despeckle))
            modified = True

    else:

        if options.auto:
            image.render(pdfPage = pdfPage)
            gray = SImage.toGray(image.get_array())
            if np.all(np.logical_or(gray == 0, gray == 255)):
                msg('auto-thresholding')
                image = PdfImage(array = gray > 127)
                modified = True

        if options.colorspace:
            image.render(pdfPage = pdfPage)
            mode = {'cmyk':'CMYK', 'rgb':'RGB', 'gray':'L', 'grey':'L'}.get(options.colorspace.lower())
            msg(f'converting colorspace: {image.get_mode()} --> {mode}')
            modified = image.change_mode(mode, intent)

        if options.bitonal and image.get_mode() != '1':
            image.render(pdfPage = pdfPage)
            cs_name = PdfColorSpace.get_name(image.ColorSpace)
            msg(f'converting {cs_name} --> 1')
            image.change_mode('L', intent)
            image = PdfImage(array = SImage.toBitonal(image.get_array(), threshold = options.threshold))
            modified = True

        if options.upsample:
            if PdfColorSpace.get_name(image.ColorSpace) == '/Indexed':
                image.render()
            msg(f'upsampling: alpha = {options.alpha}, bounds = {options.bounds}')
            image.set_array(SImage.superResolution(image.get_array(), alpha = options.alpha, bounds = options.bounds))
            modified = True

        if options.resample:
            f = options.resample
            if PdfColorSpace.get_name(image.ColorSpace) == '/Indexed':
                image.render()
            msg(f'resampling: factor = {f}, method = bicubic')
            image.set_array(SImage.resize(image.get_array(), int(image.w() * f), int(image.h() * f)))
            modified = True

        if options.zip:
            msg(f'converting to zip')
            image.set_array(image.get_array())
            modified = True

    if modified or options.colorspace and image.ColorSpace != image_obj.ColorSpace:
    
        obj = image.encode()

        image_obj.Filter = obj.Filter
        image_obj.stream = obj.stream
        image_obj.ColorSpace = obj.ColorSpace
        image_obj.BitsPerComponent = obj.BitsPerComponent
        image_obj.DecodeParms = obj.DecodeParms
        image_obj.Decode = obj.Decode
        image_obj.Width = obj.Width
        image_obj.Height = obj.Height

    # print('obj:')
    # print(obj)

def getPageRange(s:str):
    '''
    Parses a page range string formatted as 'N1[,N2-N3[,..]]' and returns a list of the form: [N1, N2, N2+1, ... N3].
    '''
    interval = lambda i: i if len(i) == 1 else [n for n in range(i[0], i[1]+1)] if len(i) == 2 else []
    return [n for r in re.split(',', options.pages) for n in interval([int(x) for x in re.split('-',r)]) ]

# ============================================================================= main()

if __name__ == '__main__':

    ap = argparse.ArgumentParser()

    ap.add_argument('inputPaths', nargs='+', metavar='FILE', help='input files: images or PDF')
    ap.add_argument('-output', '-o', type=str, metavar='PATH', help='output PDF file path')
    ap.add_argument('-pages', type=str, metavar='RANGE', help='process selected pages; RANGE = N1[,N2-N3[,..]]')
    ap.add_argument('-dpi', type=float, metavar='N', help='set resolution of input images to DPI')

    ap.add_argument('-bitonal', action='store_true', help='convert color/gray images to bitonal using Otsu\'s algorithm')
    ap.add_argument('-auto', action='store_true', help='detect if gray/color images are in fact bitonal and convert them')
    ap.add_argument('-threshold', type=int, metavar='T', default=0, help='threshold adjustment; T = -255..+255; def = 0')

    ap.add_argument('-scale2x', action='store_true', help='upsample bitonal images with the scale2x algorithm')
    ap.add_argument('-despeckle', type=int, default=0, metavar='T', help='despeckle bitonal images; T = [-12..12]; def = 0')
    ap.add_argument('-jbig2', action='store_true', help='compress bitonal images with JBIG2 (lossless)')
    ap.add_argument('-cpc', action='store_true', help='pre-process bitonal images with CPCTool before JBIG2 compression')

    ap.add_argument('-dict', type=int, default=20, metavar='N', help='pages per JBIG2 symbol dictionary; def = 20')

    ap.add_argument('-zip', action='store_true', help='compress color/gray images with JPEG')
    ap.add_argument('-jpeg', action='store_true', help='compress color/gray images with JPEG')
    ap.add_argument('-j2k', action='store_true', help='compress color/gray images with JPEG 2000')
    ap.add_argument('-quality', '-q', type=int, default=95, metavar='Q', help='JPEG/JPEG 2000 compression quality; Q=0..100 (def=95)')
    ap.add_argument('-colorspace', '-cs', type=str, metavar='S', choices=['gray', 'rgb', 'cmyk'], help='convert images to color space; S = gray|rgb|cmyk')

    ap.add_argument('-intent', type=str, metavar='I', choices = ['absolute', 'relative', 'perceptual', 'saturation', 'native', 'none'],
                    default = 'perceptual',
                    help='rendering intent; I = absolute|relative|perceptual|saturation|native|none (def=\'perceptual\')')

    ap.add_argument('-resample', type=float, metavar='F', help='resample color/gray images with bicubic interpolation')
    ap.add_argument('-upsample', action='store_true', help='upsample color/gray images x2 using Bayesian algorithm')
    ap.add_argument('-predictors', action='store_true', help='fixes incorrect PNG predictor values in images\' DecodeParms dicts')
    ap.add_argument('-alpha', type=int, metavar='A', default=1, help='alpha for the upsampling algo, higher=sharper; A=1..10, def=1')
    ap.add_argument('-bounds', type=str, metavar='B', default='softmax', choices=['softmax','local','none'], help='bounds for the upsampling algo, B=softmax|local|none, def=softmax')

    ap.add_argument('-descreen', type=float, metavar='C', help='descreen images; try confidence C = 3 (in units of stand. dev.)')

    options = ap.parse_args()

    LINE_SINGLE = '-'*64
    LINE_DOUBLE = '='*64

    intents =  {'absolute':'/AbsoluteColorimetric',
                'relative':'/RelativeColorimetric',
                'saturation':'/Saturation',
                'perceptual':'/Perceptual',
                'native':'native',
                'none':'none'}
    options.intent = intents.get(options.intent, None)
    assert options.intent != None

    fileBase, fileExt = os.path.splitext(options.inputPaths[0])

    # ---------- Input is PDF: compress and extract images ----------
    if fileExt.lower() == '.pdf':

        assert len(options.inputPaths) == 1
        pdf = PdfReader(options.inputPaths[0])
        N = len(pdf.pages)
        eprint(f"[PAGES]: {N}")

        pageRange = getPageRange(options.pages) if options.pages else [n+1 for n in range(N)]

        # JBIG2 compression
        if options.jbig2:
            cache = set()
            bitonal_images = {}
            pageCount = 0
            for pageNo in pageRange:
                print(f'[{pageNo}]')
                page = pdf.pages[pageNo-1]
                bitonal_images = bitonal_images | {id(obj):obj for name, obj in PdfObjects(page, cache=cache) 
                            if isinstance(obj, PdfDict) and obj.Subtype == PdfName.Image 
                            and (obj.BitsPerComponent == '1' or obj.ImageMask == PdfObject('true'))}
                pageCount += 1
                if pageCount == options.dict:
                    jbig2_compress(bitonal_images, cpc = options.cpc)
                    pageCount = 0
                    bitonal_images = {}
            
            if len(bitonal_images) > 0:
                jbig2_compress(bitonal_images, cpc = options.cpc)

            pdfOutPath = fileBase + ('-cpc' if options.cpc else '') + f'-jbig2' + fileExt
            print(LINE_DOUBLE)
            print(f'Writing output to {pdfOutPath}')
            PdfWriter(pdfOutPath, trailer=pdf, compress=True).write()

            sys.exit()

        # Iterate over pages
        cache = set()

        for pageNo in pageRange:

            print('-'*60)
            print(f'Processing page {pageNo}')
            print('-'*60)

            page = pdf.pages[pageNo-1]
            images = {name+f'_{id(obj)}':obj for name, obj in PdfObjects(page, cache=cache) 
                        if isinstance(obj, PdfDict) and obj.Subtype == PdfName.Image 
                        and name not in [PdfName.Mask, PdfName.SMask]}

            # print(page)
            print(f'Page {pageNo}, read {len(images)} images')
            try: defaultColorSpaces = page.Resources.ColorSpace.keys()
            except: defaultColorSpaces = None
            print(f'Page.Resources.ColorSpace:', defaultColorSpaces)
            for n,idx in enumerate(images):
                obj = images[idx]
                print(LINE_SINGLE)
                print(f'[{n+1}] Image ID = {idx}')
                pprint(obj)
                print(LINE_SINGLE)

                # ---------- Modify images ----------
                if options.zip or options.upsample or options.resample or options.colorspace \
                        or options.bitonal or options.predictors \
                        or options.despeckle \
                        or options.scale2x \
                        or options.auto:

                    pageArg = page if options.colorspace else None
                    modify_image_xobject(obj, pageArg, options)
                    print(LINE_SINGLE)
                    print("RESULT:")
                    pprint(obj)

                # ---------- Recompress images ----------
                elif options.jpeg or options.j2k:
                    if PdfImage.xobject_get_bpc(obj) != 1:
                        msg('recompressing image')
                        PdfImage.recompress(obj, Format='JPEG' if options.jpeg else 'JPEG2000' , Q=options.quality)
                        # if image_obj.SMask:
                        #     msg('recompressing mask')
                        #     PdfImage.recompress(image_obj.SMask, Format='JPEG', Q=options.quality)

                # ---------- Descreen images ----------
                elif options.descreen:
                    if PdfImage.xobject_get_bpc(obj) != 1:
                        msg('descreening image')
                        image = PdfImage(obj = obj)
                        image.set_array(SImage.descreen(image.get_array(), options.descreen))
                        obj_new = image.encode()
                        PdfImage.xobject_copy(obj_new, obj)

                # ---------- Extract images ----------
                else:

                    # image = PdfImage.decode(obj,page, adjustColors = False, applyMasks = True, applyIntent = options.applyIntent)
                    image = PdfImage(obj=obj)
                    if options.dpi:
                        image.dpi = (options.dpi, options.dpi)
                    stream, ext = image.saveAs('auto', render = True)
                    outputPath = fileBase  +f'.page{pageNo}.image{n+1}' + ext
                    open(outputPath, 'wb').write(stream)

        if options.zip or options.jpeg or options.j2k or options.upsample or options.resample or options.colorspace \
                or options.bitonal or options.predictors or options.descreen \
                or options.despeckle or options.scale2x or options.auto:
            suffix = ''
            if options.auto: suffix += f'-auto'
            if options.despeckle: suffix += f'-despeckle={options.despeckle}'
            if options.scale2x: suffix += f'-scale2x'
            if options.upsample: suffix += f'-upsample={options.alpha}'
            if options.resample: suffix += '-resample'
            if options.colorspace: suffix += f'-{options.colorspace}'
            if options.zip: suffix += f'-zip'
            if options.jpeg: suffix += f'-jpeg={options.quality}'
            if options.j2k: suffix += f'-j2k={options.quality}'
            if options.bitonal: suffix += '-bitonal'
            if options.predictors: suffix += '-predictors'
            if options.descreen: suffix += f'-descreen={options.descreen}'
            assert suffix != ''
            pdfOutPath = fileBase + suffix + fileExt
            print(LINE_DOUBLE)
            if (options.j2k): print('Warning: *** JPEG 2000 output may be buggy; please check manually ***')
            print(f'Writing output to {pdfOutPath}')
            PdfWriter(pdfOutPath, trailer=pdf, compress=True).write()

    # ---------- Input is images: convert to PDF ----------
    else:

        pdfPath = options.output or options.inputPaths[0]+".pdf"

        pdf = PdfWriter(pdfPath,compress=True)
        N = len(options.inputPaths)
        for n,imagePath in enumerate(options.inputPaths):
            base,ext = os.path.splitext(imagePath)
            if ext.lower() == '.jp2':
                image = PdfImage(jp2 = open(imagePath,'rb').read())
            else:
                image = PdfImage(pil = Image.open(imagePath))

            print(f'[{n+1}/{N}] {image}')
            print(f'[{n+1}/{N}] {imagePath}')

            image.dpi = (options.dpi, options.dpi) if options.dpi != None \
                    else image.dpi if image.dpi not in [None, (1,1)] \
                    else (72,72)

            page = image.to_pdf_page()
            pdf.addPage(page)

        pdf.write()
        print(f'Output written to {pdfPath}')