Jpeg2000Encoder Class

Jpeg2000BlockCoder.HighThroughput codes each code-block with one HT cleanup pass of ISO/IEC 15444-15 (no HT refinement pass), and declares it in the codestream (the CAP marker segment, the code-block style); a Jpeg2000Container.Jp2 output is then a JPH file, of the brand jph . The pass codes the same quantized coefficients as the standard coder: a lossless output is the image, and a lossy output decodes to the samples of the standard coder at the same Jpeg2000Encoder.Quality.

The pass has no fractional bit-planes and cannot be split between layers: the output has one quality layer, and Jpeg2000Encoder.QualityLayers may only hold one layer. Its target is reached by choosing, for each code-block, the lowest bit-plane its pass codes (each candidate is coded, at the same rate-distortion slopes as the coding passes of the standard coder), so that a target is reached in steps of whole bit-planes of the code-blocks.

With the Jpeg2000BlockCoder.HighThroughput block coder, the file is a JPH file.

Each level halves the size of the lowest resolution: more levels compress large smooth images slightly better and let a reader decode smaller thumbnails (a decode at 1/2^n of the size discards n levels).

Without precincts, a resolution level of a tile-component is one position: Jpeg2000ProgressionOrder.Pcrl and Jpeg2000ProgressionOrder.Cprl then give the same packets, component by component, and Jpeg2000ProgressionOrder.Rpcl the packets of Jpeg2000ProgressionOrder.Rlcp with the layers of a component together.

The quality sets the quantization step sizes of the sub-bands; there is no rate control, so the size of the output follows from the image. With the scale of JpegEncoder.Quality, s = 5000 / Quality below 50 and s = 200 - 2 * Quality otherwise (at least 1), the error that the quantization of a sub-band of the luminance (or of the gray levels) adds to the image has a step of 0.47 * s^(3/4) levels of an 8-bit sample, proportionally more for deeper samples; the two chrominances are quantized twice as coarsely.

The values are calibrated on JpegEncoder: on photographs, a quality gives the structural similarity (SSIM) of a JPEG file of the same quality, with chrominances that are not worse than its sub-sampled ones, in a file of about two thirds of the size (half of it at the quality 20). 100 is the finest quantization, still lossy.

Each layer adds coding passes to the previous ones, so that a reader that stops after a layer gets the best image of that size: with the Jpeg2000ProgressionOrder.Lrcp order, the bytes of the first layers are the start of the output. A target of a size or of a number of bits per pixel is reached by truncating the coding passes of the code-blocks at the points that reduce the squared error of the image most for their bytes (the post-compression rate-distortion optimization of T.800, J.14); a layer whose target is not above the size of the previous ones adds nothing. A target only removes coding passes: Jpeg2000Encoder.Quality still sets the finest quantization of a lossy output, and a lossless output whose last layer has a target is not lossless when the target cuts coding passes.

The coded bytes of every code-block are kept until the layers are chosen, as with one layer; choosing the layers writes the packet headers a few dozen times. Jpeg2000QualityLayer.Complete is only allowed last.

Each tile is transformed and coded on its own, so that the encoder holds the coefficients of one tile at a time instead of those of the whole image: a large image is coded with tiles of 512 to 2048 pixels without holding its transformation. A reader can also decode one tile alone. Small tiles cost bytes and quality at their edges. An image has at most 65,535 tiles.