Merge branch 'feature/optimization-fixes'

docs/introduction.rst

@@ -112,7 +112,7 @@ Web front-ends
 
 The Docker image ``ocrmypdf-alpine`` provides a web service front-end that allows files to submitted over HTTP and the results "downloaded". This is an HTTP server intended to simplify web services deployments; it is not intended to be deployed on the public internet and no real security measures to speak of.
 
-In addition, the following integrations are available:
+In addition, the following third-party integrations are available:
 
 * `Nextcloud OCR <https://github.com/janis91/ocr>`_ is a free software plugin for the Nextcloud private cloud software
 

src/ocrmypdf/exec/pngquant.py

@@ -17,6 +17,9 @@
 
 from functools import lru_cache
 from subprocess import run
+from tempfile import NamedTemporaryFile
+
+from PIL import Image
 
 from . import get_version
 from ..exceptions import MissingDependencyError
@@ -36,16 +39,32 @@ def available():
 
 
 def quantize(input_file, output_file, quality_min, quality_max):
-    args = [
-        'pngquant',
-        '--force',
-        '--skip-if-larger',
-        '--output',
-        output_file,
-        '--quality',
-        f'{quality_min}-{quality_max}',
-        '--',
-        input_file,
-    ]
-    proc = run(args)
-    proc.check_returncode()
+    if input_file.endswith('.jpg'):
+        im = Image.open(input_file)
+        with NamedTemporaryFile(suffix='.png') as tmp:
+            im.save(tmp)
+            args = [
+                'pngquant',
+                '--force',
+                '--skip-if-larger',
+                '--output',
+                output_file,
+                '--quality',
+                f'{quality_min}-{quality_max}',
+                '--',
+                tmp.name,
+            ]
+            run(args)
+    else:
+        args = [
+            'pngquant',
+            '--force',
+            '--skip-if-larger',
+            '--output',
+            output_file,
+            '--quality',
+            f'{quality_min}-{quality_max}',
+            '--',
+            input_file,
+        ]
+        run(args)

src/ocrmypdf/optimize.py

@@ -25,7 +25,7 @@ from pathlib import Path
 from PIL import Image
 
 import pikepdf
-from pikepdf import Name, Dictionary
+from pikepdf import Name, Dictionary, Array
 
 from . import leptonica
 from ._jobcontext import JobContext
@@ -149,11 +149,22 @@ def extract_image_generic(*, pike, root, log, image, xref, options):
 def extract_images(pike, root, log, options, extract_fn):
     """Extract image using extract_fn
 
-    extract_fn decides whether the image is interesting in this case
+    Enumerate images on each page, lookup their xref/ID number in the PDF.
+    Exclude images that are soft masks (i.e. alpha transparency related).
+    Record the page number on which an image is first used, since images may be
+    used on multiple pages (or multiple times on the same page).
+
+    Current we do not check Form XObjects or other objects that may contain
+    images, and we don't evaluate alternate images or thumbnails.
+
+    extract_fn must decide if wants to extract the image in this context. If
+    it does a tuple should be returned: (xref, ext) where .ext is the file
+    extension. extract_fn must also extract the file it finds interesting.
     """
 
     include_xrefs = set()
     exclude_xrefs = set()
+    pageno_for_xref = {}
     errors = 0
     for pageno, page in enumerate(pike.pages):
         try:
@@ -169,6 +180,8 @@ def extract_images(pike, root, log, options, extract_fn):
                 smask_xref = image.SMask.objgen[0]
                 exclude_xrefs.add(smask_xref)
             include_xrefs.add(xref)
+            if xref not in pageno_for_xref:
+                pageno_for_xref[xref] = pageno
 
     working_xrefs = include_xrefs - exclude_xrefs
     for xref in working_xrefs:
@@ -178,13 +191,12 @@ def extract_images(pike, root, log, options, extract_fn):
                 pike=pike, root=root, log=log, image=image, xref=xref, options=options
             )
         except Exception as e:
-            log.debug("Image xref %s", xref)
-            log.debug(repr(e))
+            log.debug("Image xref %s, error %s", xref, repr(e))
             errors += 1
         else:
             if result:
                 _, ext = result
-                yield pageno, xref, ext
+                yield pageno_for_xref[xref], xref, ext
 
 
 def extract_images_generic(pike, root, log, options):
@@ -198,7 +210,7 @@ def extract_images_generic(pike, root, log, options):
             pngs.append(xref)
         elif ext == '.jpg':
             jpegs.append(xref)
-    log.debug("Optimizable images: " "JPEGs: %s PNGs: %s", len(jpegs), len(pngs))
+    log.debug("Optimizable images: JPEGs: %s PNGs: %s", len(jpegs), len(pngs))
     return jpegs, pngs
 
 
@@ -216,7 +228,7 @@ def extract_images_jbig2(pike, root, log, options):
     jbig2_groups = {
         group: xrefs for group, xrefs in jbig2_groups.items() if len(xrefs) > 0
     }
-    log.debug("Optimizable images: " "JBIG2 groups: %s", (len(jbig2_groups),))
+    log.debug("Optimizable images: JBIG2 groups: %s", (len(jbig2_groups),))
     return jbig2_groups
 
 
@@ -271,7 +283,7 @@ def convert_to_jbig2(pike, jbig2_groups, root, log, options):
     must be lossy encoding since jbig2enc does not support refinement coding.
 
     When the JBIG2 symbolic coder is not used, each JBIG2 stands on its own
-    and needs no dictionary. Currently this is must be lossless JBIG2.
+    and needs no dictionary. Currently this must be lossless JBIG2.
     """
 
     _produce_jbig2_images(jbig2_groups, root, log, options)
@@ -309,7 +321,7 @@ def transcode_jpegs(pike, jpegs, root, log, options):
         # https://github.com/python-pillow/Pillow/issues/1144
         with Image.open(fspath(in_jpg)) as im:
             im.save(fspath(opt_jpg), optimize=True, quality=options.jpeg_quality)
-        # pylint: disable=no-member
+
         if opt_jpg.stat().st_size > in_jpg.stat().st_size:
             log.debug("xref %s, jpeg, made larger - skip", xref)
             continue
@@ -319,7 +331,7 @@ def transcode_jpegs(pike, jpegs, root, log, options):
         im_obj.write(compdata.read(), filter=Name.DCTDecode)
 
 
-def transcode_pngs(pike, pngs, root, log, options):
+def transcode_pngs(pike, images, image_name_fn, root, log, options):
     if options.optimize >= 2:
         png_quality = (
             max(10, options.png_quality - 10),
@@ -328,44 +340,64 @@ def transcode_pngs(pike, pngs, root, log, options):
         with concurrent.futures.ThreadPoolExecutor(
             max_workers=options.jobs
         ) as executor:
-            for xref in pngs:
+            for xref in images:
+                log.info(image_name_fn(root, xref))
                 executor.submit(
                     pngquant.quantize,
-                    png_name(root, xref),
+                    image_name_fn(root, xref),
                     png_name(root, xref),
                     png_quality[0],
                     png_quality[1],
                 )
 
-    for xref in pngs:
+    for xref in images:
         im_obj = pike.get_object(xref, 0)
-
-        # Open, transcode (!), package for PDF
         try:
-            pix = leptonica.Pix.open(png_name(root, xref))
-            if pix.depth == 1:
-                pix = pix.invert()  # PDF assumes 1 is black for monochrome
-            compdata = pix.generate_pdf_ci_data(leptonica.lept.L_FLATE_ENCODE, 0)
+            compdata = leptonica.CompressedData.open(png_name(root, xref))
         except leptonica.LeptonicaError as e:
+            # Most likely this means file not found, i.e. quantize did not
+            # produce an improved version
             log.error(e)
             continue
 
-        # This is what we should be doing: open the compressed data without
-        # transcoding. However this shifts each pixel row by one for some
-        # reason.
-        # compdata = leptonica.CompressedData.open(png_name(root, xref))
+        # If re-coded image is larger don't use it - we test here because
+        # pngquant knows the size of the temporary output file but not the actual
+        # object in the PDF
         if len(compdata) > int(im_obj.stream_dict.Length):
-            continue  # If we produced a larger image, don't use
+            log.debug(
+                f"pngquant: pngquant did not improve over original image "
+                f"{len(compdata)} > {int(im_obj.stream_dict.Length)}"
+            )
+            continue
 
-        predictor = None
-        if compdata.predictor > 0:
-            predictor = Dictionary(Predictor=compdata.predictor)
+        # When a PNG is inserted into a PDF, we more or less copy the IDAT section from
+        # the PDF and transfer the rest of the PNG headers to PDF image metadata.
+        # One thing we have to do is tell the PDF reader whether a predictor was used
+        # on the image before Flate encoding. (Typically one is.)
+        # According to Leptonica source, PDF readers don't actually need us
+        # to specify the correct predictor, they just need a value of either:
+        #   1 - no predictor
+        #   10-14 - there is a predictor
+        # Leptonica's compdata->predictor only tells TRUE or FALSE
+        # From there the PNG decoder can infer the rest from the file.
+        # In practice the predictor should be Paeth, 14, so we'll use that.
+        # See:
+        #   - PDF RM 7.4.4.4 Table 10
+        #   - https://github.com/DanBloomberg/leptonica/blob/master/src/pdfio2.c#L757
+        predictor = 14 if compdata.predictor > 0 else 1
+        dparms = Dictionary(Predictor=predictor)
+        if predictor > 1:
+            dparms.BitsPerComponent = compdata.bps  # Yes, this is redundant
+            dparms.Colors = compdata.spp
+            dparms.Columns = compdata.w
 
         im_obj.BitsPerComponent = compdata.bps
         im_obj.Width = compdata.w
         im_obj.Height = compdata.h
 
         if compdata.ncolors > 0:
+            # .ncolors is the number of colors in the palette, not the number of
+            # colors used in a true color image
             palette_pdf_string = compdata.get_palette_pdf_string()
             palette_data = pikepdf.Object.parse(palette_pdf_string)
             palette_stream = pikepdf.Stream(pike, bytes(palette_data))
@@ -378,13 +410,17 @@ def transcode_pngs(pike, pngs, root, log, options):
             cs = palette
         else:
             if compdata.spp == 1:
-                cs = Name.DeviceGray
+                # PDF interprets binary-1 as black in 1bpp, but PNG sets
+                # black to 0 for 1bpp. Create a palette that informs the PDF
+                # of the mapping.
+                palette = [Name.Indexed, Name.DeviceGray, 1, b"\xff\x00"]
+                cs = palette
             elif compdata.spp == 3:
                 cs = Name.DeviceRGB
             elif compdata.spp == 4:
                 cs = Name.DeviceCMYK
         im_obj.ColorSpace = cs
-        im_obj.write(compdata.read(), filter=Name.FlateDecode, decode_parms=predictor)
+        im_obj.write(compdata.read(), filter=Name.FlateDecode, decode_parms=dparms)
 
 
 def optimize(input_file, output_file, log, context):
@@ -404,11 +440,14 @@ def optimize(input_file, output_file, log, context):
     pike = pikepdf.Pdf.open(input_file)
 
     root = Path(output_file).parent / 'images'
-    root.mkdir(exist_ok=True)  # pylint: disable=no-member
+    root.mkdir(exist_ok=True)
 
     jpegs, pngs = extract_images_generic(pike, root, log, options)
     transcode_jpegs(pike, jpegs, root, log, options)
-    transcode_pngs(pike, pngs, root, log, options)
+    # if options.optimize >= 2:
+    # Try pngifying the jpegs
+    #    transcode_pngs(pike, jpegs, jpg_name, root, log, options)
+    transcode_pngs(pike, pngs, png_name, root, log, options)
 
     jbig2_groups = extract_images_jbig2(pike, root, log, options)
     convert_to_jbig2(pike, jbig2_groups, root, log, options)