Module jpeg 

A JPEG decoder that reproduces libjpeg(-turbo)’s output byte for byte — what pdfium’s DCTDecode produces (core/fxcodec/jpeg, libjpeg at its defaults: the JDCT_ISLOW integer IDCT, “fancy” triangle-filter chroma upsampling, the 16.16 fixed-point YCbCr→RGB tables of jdcolor.c).

Neither pure-Rust decoder in the dependency tree matches those bytes: zune-jpeg (the image crate’s) rounds its upsampler and colour conversion differently, jpeg-decoder carries stb_image’s +8/+8 h2v2 filter where libjpeg alternates +8/+7 and a float-derived colour conversion. The pixel differences are ±1, invisible — and exactly what a byte-for-byte oracle against pdfium’s raster cannot tolerate. So this is a deliberately small decoder: baseline and progressive Huffman, 8-bit, 1 or 3 components, restart intervals, and libjpeg’s 1/2, 1/4, 1/8 DCT-scaled output (jidctred.c’s 4×4/2×2/1×1 transforms with jdmaster.c’s per-component scaling and the scaled upsampler rules); everything else (arithmetic coding, 12-bit, lossless, CMYK/YCCK, DNL) is reported as unsupported and the page stays with pdfium. Checked against Pillow’s libjpeg-turbo on docling-pdf’s tests/data/jpeg/ fixtures (raster::jpeg_tests::matches_libjpeg_on_the_fixtures).

It lives in docling-core because Pillow decodes with the same libjpeg: besides docling-pdf’s raster/renderer (which re-export it), the DocLang serializer hashes a JPEG picture’s decoded pixels to name its asset the way docling does (pixel_digest.rs).

Structs

Image

A decoded image: channels is 1 (gray) or 3 (RGB, or the raw component triplets when no colour transform applies), rows of width pixels.

Info

Header facts the caller needs before decoding.

Enums

Error

Functions

decode

Decode data at 1/scale_denom (1, 2, 4 or 8 — libjpeg’s DCT scaling, which pdfium requests for an image at least twice the bitmap’s size: resolution_levels_to_skip); the output is ceil(w/scale_denom) × ceil(h/scale_denom). color_transform is the PDF’s /ColorTransform decode parameter (default 1), which pdfium forces on when an Adobe marker is present and otherwise uses to decide whether a 3-component image is converted from YCbCr or handed over raw.

info

Read the frame header only.