Rate–Distortion Optimised Image Compression with Directional Intra Prediction, Adaptive Transforms and Context-Adaptive Arithmetic Coding

Abstract

This paper presents a complete still-image codec built onrate–distortion (RD) optimised block coding. The design combines aquadtree partition over 16×16 to 4×4 coding units, thirty-five directionalintra prediction modes with most-probable-mode signalling, adaptiveselection between the DCT-II and the DST-VII, trellis quantisation withexplicit last-position optimisation, and a context-adaptive binaryarithmetic coder with approximately 250 adaptive contexts. Rate controlis driven by an activity-adaptive Lagrangian multiplier. On the Kodaktest set the codec achieves a Bjøntegaard delta-rate of −36.8 % againstJPEG in PSNR, level with WebP at method 6 (−36.4 % on the samemeasurement), and 6.4 points ahead of WebP in SSIM. Encoding isparallelised by a wavefront over coding units that is bit-exact with theserial encoder, and a set of rate–distortion search prunings bringsencoding time to within a few per cent of WebP at method 6 for a loss of0.11 dB, while cross-component chroma prediction, independentlydecodable slices and 32×32 coding units close most of the remaininggap. An optional archival mode for scanned documents lifts the 4:2:0fidelity ceiling and leads both anchors by 4.1 dB at 6 bpp. We furtherreport three negative results of independent interest: for the separableDCT dictionaries formed by Kronecker products that motivated anearlier version of this work, Orthogonal Matching Pursuit is provablyand empirically identical to simple coefficient thresholding, and thesparse-coding machinery it requires yields no coding gain while costingan order of magnitude in encoding time.