j2k 1.0.0

Pure Dart JPEG 2000 codec: decodes JP2 and J2K files to 8- or 16-bit pixels with alpha, encodes pixel buffers and PGM/PPM, and runs on the VM, dart2js and dart2wasm.

1.0.0

First stable release. The public surface — decodeJpeg2000, decodeJpeg2000Source, probeJpeg2000, encodeJpeg2000Pixels, encodeJpeg2000, encodeJpeg2000Source, the option classes and the sealed Jpeg2000Exception hierarchy — is unchanged from 0.9.0 and is now covered by semantic versioning. Everything below is additive or a fix.

Upgrade from 0.9.0 is not optional on the web

0.9.0 fails on every 8-bit non-RGB decode compiled with dart2js: gray, gray+alpha, RGBA and multi-component images all throw. pixels16![index] inside the sample loop let dart2js hoist the receiver null check out of the loop, so the null 16-bit buffer was dereferenced on paths that never touch it. The Dart VM and dart2wasm were unaffected, which is why it survived release. If you ship 0.9.0 to a browser, upgrade.

Added

• decodeJpeg2000Parallel and Jpeg2000ParallelOptions: decode spread over worker isolates, behind if (dart.library.io), with the sequential path as the default and as the web fallback. Splits by tile, or by code block when there are fewer tiles than workers — the single-tile image typical of PDF. The code-block path drives the real StdEntropyDecoder, so its output is bit-identical to the sequential decode by construction. Measured on eight cores: 2.6x to 3.8x above 0.2 megapixel; below minParallelPixels (36864 samples, measured not guessed) it stays sequential. decodeJpeg2000 remains synchronous and untouched.
• Subsampled components in raw codestreams (4:2:0, 4:2:2) now decode instead of throwing. Resampling moved out of Resampler, which needed a ColorSpace a raw codestream does not have, into a colour-free ComponentUpsampler; Resampler became a thin adapter, so the package has one resampling implementation rather than two. Subsampling factors now span 1 to 255, the full range of XRsiz/YRsiz in ISO/IEC 15444-1 A.5.1, instead of being capped at 2:1.
• ATK marker (ISO/IEC 15444-2): wavelet kernels signalled by ATK index are resolved and synthesised by the generic lifting procedure of Annex G. Installation of the filter is deferred to the end of each header because an ATK may follow the COD that uses it. Validation: the 5-3 and 9-7 kernels described through ATK reproduce the specialised filters bit for bit; the specialised filters were not replaced. Still rejected, with the clause in the message: the arbitrary category of Annex H and 128-bit coefficient types. The encoder remains Part 1.
• Encoder: per-component bit depth and signed samples; the multiple component transform is switched off automatically when the first three components differ in depth or subsampling, as G.2 requires.
• Float32x4 (SIMD) synthesis for the irreversible path: 9/7 inverse wavelet, inverse ICT and RGB pixel collection, bit-exact against the JJ2000 fixtures. A 2717x3701 RGB image went from 1236 to 417 ms on the VM JIT and 1436 to 493 ms in AOT.

Fixed

• dart2js: every 8-bit non-RGB decode threw. See above.
• Jpeg2000DecodeOptions.resolution threw RangeError for any level below full resolution. The inverse wavelet adapter answered tile, component and corner geometry at the subband-tree level (always the total number of decomposition levels) while image width and height answered at the target level; the two coincide only at full resolution. Sizes now follow the B.5 formula at the target level, and synthesis allocates the reduced buffer instead of allocating the full one and using a corner. Full-resolution decodes are byte-identical to before. This is a deliberate divergence from the JJ2000 reference, which carries the same defect; it is documented in doc/DIVERGENCIAS_JJ2000.md so nobody "fixes" it back.
• resolution was documented backwards, as the number of levels to discard. It is the level to reconstruct: 0 is the coarsest (the LL band alone) and NL is the full image. A level above the maximum is clamped with a warning.
• getCompImgWidth returned compW * subs, which overruns the image edge when the width is not a multiple of the subsampling factor — a 4:2:0 JP2 nine pixels wide decoded wrong. Inherited from JJ2000.
• The progression order in a POC marker was not validated, unlike COD and COC; an out-of-range Ppoc reached the default branch. It is now rejected with the value that was read, instead of being reported as an unsupported feature.

0.9.0

First release, published as j2k because the name jpeg2000 was already taken on pub.dev. The API is the same: decodeJpeg2000, probeJpeg2000, encodeJpeg2000Pixels, encodeJpeg2000.

• JP2 and raw J2K decoder ported from JJ2000: codestream parsing, EBCOT/MQ entropy decoding, ROI de-scaling, dequantization, 5x3 and 9x7 inverse wavelets, inverse RCT/ICT, and JP2 colour handling (enumerated colour spaces, restricted ICC profiles, palettes, channel definitions). Bit-exact against the JJ2000 reference on the bundled conformance subset.
• decodeJpeg2000 returns interleaved pixels as gray, gray+alpha, RGB, RGBA or raw multi-component data, 8 bits per sample by default or 16 with outputBitDepth: 16, with hasAlpha, alphaIsPremultiplied and the source bit depths.
• probeJpeg2000 reads geometry, bit depths, tiling and alpha from the headers without decoding.
• Jpeg2000DecodeOptions.maxPixels and maxDimension reject oversized images before any allocation; onWarning receives non-fatal diagnostics.
• Sealed Jpeg2000Exception hierarchy: format, truncated, corrupted, unsupported and budget errors.
• encodeJpeg2000Pixels encodes interleaved pixel buffers, 1 to 16 bits per sample, with alpha recorded in the JP2 channel definition box; encodeJpeg2000 takes binary PGM (P5) and PPM (P6) bytes, 8 or 16 bits. Both produce raw J2K or JP2, lossless or rate-controlled, with optional tiling.
• Byte-oriented API with no dart:io in the import graph; works on the Dart VM, dart2js and dart2wasm. decodeJpeg2000Source and encodeJpeg2000Source accept files and paths on the VM and Blob/File in browsers.
• Command-line tools jp2dec and jp2enc.