JpegEncoder Class

Subsampled chroma is the mean of the full-resolution samples each chroma sample covers (box filter, centered siting). The setting has no effect on gray pixel formats, which are written as a single component.

The image is coded twice from its quantized coefficients: once to count the symbols of each table, from which length-limited codes are built (ITU-T T.81 section K.2: at most 16 bits, no code made of 1-bits only), then to write the scan with them. The quantized coefficients, and so the decoded pixels, are those of the default output; only the DHT segment and the size of the entropy-coded data change, usually a few percent smaller. As with the progressive process, the coefficients of the whole image are kept in memory until the scan is written (two bytes per coefficient), where the default baseline output only keeps a few rows. It has no effect on JpegEncodingProcess.Progressive, whose scans always use tables computed from their own symbols.

Both processes store the same quantized coefficients, so they decode to the same pixels. JpegEncodingProcess.Progressive writes them in several scans (six for a grayscale image, ten for a color one), each with Huffman tables computed from its own statistics: the DC coefficients without their lowest bit, the first five AC coefficients of the luminance without their two lowest bits, the chrominance without its lowest bit, the rest of the luminance, then the missing bits. The coefficients of the whole image are kept in memory until the last scan (two bytes per coefficient), where the baseline process only keeps a few rows.

The Annex K tables are scaled linearly: with s = 5000 / Quality below 50 and s = 200 - 2 * Quality otherwise, each quantization value is clamp((base * s + 50) / 100, 1, 255). Quality 50 uses the Annex K tables unchanged and 100 quantizes every coefficient by 1 (still lossy: color conversion and rounding remain).

With a nonzero interval, a DRI segment precedes the first scan and each scan is split into intervals of that many MCUs, in raster order (ITU-T T.81 sections B.2.4.4 and F.1.2.3): every interval but the last ends with the 1-bits that complete its last byte and a marker from RST0 to RST7 (cyclically), and the next interval starts again with DC predictions of 0 (and, in a progressive scan, without a pending end-of-band run). A decoder can then resynchronize after corrupted data, or decode the intervals independently. The decoded pixels do not change: only a few bytes are added per marker, and the end-of-band runs of a progressive scan are shorter.

An MCU of the interleaved scans covers 8x8 pixels for a grayscale image or 4:4:4 chroma, 16x8 for 4:2:2 and 16x16 for 4:2:0; in the single-component scans of the progressive process, it is one 8x8 block of the component. An interval equal to the number of MCUs of a row of the baseline scan, ceil(width / MCU width), puts a marker before every MCU row but the first.