Fix handling of DPI for rare case of JPEG recompression after deskew/clean

This test is exercised by page 4 of multipage.pdf. If all images are
JPEGs, and one of deskew/clean removes DPI information, make sure that
we can get the right information back and that the DPI stays square.

RELEASE_NOTES.rst

@@ -9,6 +9,7 @@ Download software here: https://github.com/jbarlow83/OCRmyPDF/tags
 v4.2 (planned):
 ===============
 
+-  Improved support for PDFs containing images with "non-square" pixel aspect ratios, such as 200x100 DPI.
 -  Support for some older programs dropped - should not be a problem for anyone except Ubuntu 12.04 LTS users
 
    +  Ghostscript 9.15 is now required
@@ -16,7 +17,6 @@ v4.2 (planned):
 
 -  Ghostscript now runs in "safer" mode where possible
 
-
 v4.1.4:
 =======
 

ocrmypdf/main.py

@@ -738,12 +738,20 @@ def select_image_for_pdf(
     image = next(ii for ii in infiles if ii.endswith(image_suffix))
 
     pageinfo = get_pageinfo(image, pdfinfo, pdfinfo_lock)
-    if all(image['enc'] == 'jpeg' for image in pageinfo['images']):
+    if all(orig_image['enc'] == 'jpeg' for orig_image in pageinfo['images']):
         # If all images were JPEGs originally, produce a JPEG as output
         im = Image.open(image)
-        fallback_dpi = get_page_dpi(pageinfo)
-        dpi = im.info.get('dpi', fallback_dpi)
-        dpi = round(dpi[0]), round(dpi[1])  # Pillow requires integer DPI
+
+        # At this point the image should be a .png, but deskew, unpaper might
+        # have removed the DPI information. In this case, fall back to square
+        # DPI used to rasterize. When the preview image was rasterized, it
+        # was also converted to square resolution, which is what we want to
+        # give tesseract, so keep it square.
+        fallback_dpi = get_page_square_dpi(pageinfo)
+        dpi = im.info.get('dpi', (fallback_dpi, fallback_dpi))
+
+        # Pillow requires integer DPI
+        dpi = round(dpi[0]), round(dpi[1])
         im.save(output_file, format='JPEG', dpi=dpi)
     else:
         re_symlink(image, output_file)
@@ -871,7 +879,7 @@ def add_text_layer(
     y2 = page_image.mediaBox.getUpperRight_y()
 
     # Rotation occurs about the page's (0, 0). Most pages will have the media
-    # box at (0, 0) will all content in the first quadrant but some cropped
+    # box at (0, 0) with all content in the first quadrant but some cropped
     # files may have an offset mediabox. We translate the page so that its
     # bottom left corner after rotation is pinned to (0, 0) with the image
     # in the first quadrant.

tests/resources/README.rst

@@ -14,9 +14,9 @@ In some cases they were converted from one image format to another without other
 +=====================+================================================================================+
 | c02-22.pdf          | `Project Gutenberg`_, Adventures of Huckleberry Finn, page 22                  |
 +---------------------+--------------------------------------------------------------------------------+
-| congress.jpg        | `US Congressional Records`_                                                    |
+| congress.jpg        | `US Congressional Records`_ (Public Domain)                                    |
 +---------------------+--------------------------------------------------------------------------------+
-| graph.pdf           | `Wikimedia: Pandas text analysis.png`_                                         |
+| graph.pdf           | `Wikimedia: Pandas text analysis.png`_ (Public Domain)                         |
 +---------------------+--------------------------------------------------------------------------------+
 | lichtenstein.pdf    | `Wikimedia: JPEG2000 Lichtenstein`_ (Creative Commons BY-SA 3.0)               |
 +---------------------+--------------------------------------------------------------------------------+

tests/test_main.py

@@ -565,8 +565,8 @@ def test_algo4():
 
 
 @pytest.mark.parametrize('renderer', [
-    'hocr'])
-def test_non_square_resolution(renderer):
+    'hocr'])  # tesseract cannot pass this test yet
+def test_non_square_resolution(renderer, spoof_tesseract_cache):
     # Confirm input image is non-square resolution
     in_pageinfo = pdf_get_all_pageinfo(_infile('aspect.pdf'))
     assert in_pageinfo[0]['xres'] != in_pageinfo[0]['yres']
@@ -574,7 +574,7 @@ def test_non_square_resolution(renderer):
     out = 'aspect_%s.pdf' % renderer
     check_ocrmypdf(
         'aspect.pdf', out,
-        '--pdf-renderer', renderer)
+        '--pdf-renderer', renderer, env=spoof_tesseract_cache)
 
     out_pageinfo = pdf_get_all_pageinfo(_outfile(out))