Technical Report No. 99, July 1996 - Abstract

Hartenstein, Hannes; Wu, Xiaolin:
Analysis of Trellis Quantization for Near-Lossless Image Coding

In this paper we discuss near-lossless image compression by trellis quantization (TQ) coupled with predictive coding. Here near-lossless coding means-constrained coding, i.e., each pixel can be altered by at most grey values, where is a pregiven threshold. In a predictive coding scheme the simplest way to guarantee the bound is to perform a uniform quantization of the prediction residues with quantizer bin size matching the required tolerance. The near-lossless version of CALIC uses this mechanism. But since quantization is inside the DPCM loop changing a pixel value at the current position affects subsequent predictions of forthcoming pixels. The trellis quantization scheme proposed by Ke and Marcellin tries to take into account those global implications of quantization. However, despite its high computational cost the TQ scheme is inferior to the near-lossless CALIC. In this paper we reexamine the trellis quantization approach for near-lossless image coding, and point out its weaknesses that render TQ ineffective. We also incorporate into TQ context modeling techniques inspired from CALIC. We report coding results for various standard test images.