Planetary Image Compression By Trellis Coded Quantization

Abstract

Image compression schemes for deep-space communication are compared in terms of their complexity and performance. The comparison includes Voyager’s entropy coded DPCM, an entropy coded DCT-based scheme, and a DCT-based scheme with trellis coded quantization. While the performance of the latter scheme is slightly inferior, its low complexity could be appealing for on-board encoders. 1. Introduction. The image compression scheme used on the Voyager spacecraft was constrained by the need to use 1970’s space-qualified hardware and essentially consisted of Huffman coding applied to the differences between adjacent pixels (entropy coded DPCM) [l]. With the availability of enormously more powerful VLSI hardware in the 1990s and beyond, it is now possible to envision a future in which source coding can deliver gains to deep-space telemetry fully comparable to those already realized by channel coding. One of the possible strategies for efficient image compression is the adoption of the emerging JPEG standard [2] for still-image compression. This method is based on performing a discrete cosine transform (DCT) on 8 x 8 blocks of the picture, quantizing the resulting coefficients, and then applying Huffman coding or arithmetic coding. Its performance has been used as a reference, together with the results achieved by the present Voyager’s compression system. These reference results show that the JPEG standard can achieve a compression ratio of 1O:l with an average degradation of only 1 out of 256 gray levels on a typical planetary image. Deep-space applications require improvements in three areas: low on-board complexity, high ratedistortion performance, 2nd low sensitivity to channel errors. The first requirement was addressed by developing an integer cosine transform (ICT) which reduces the complexity by requiring only integer operations on small integers and at the same time gives a rate-distortion performance very close to that offered by the floating-point DCT. Performance comparisons on the compression of planetary images are given for DCT and ICT based schemes. The second and third requirements were addressed by studying the possible use of trellis coded quantization (TCQ), which was recently developed by Marcellin and Fischer [3] for effcient encoding of memoryless and Gauss-Markov sources with moderate complexity implementation. TCQ replaces the scalar quantization and entropy coding in the JPEG scheme. The motivation for TCQ lies in Ungerboeck’s work on trellis coded modulation (TCM) and in the theoretical results of alphabet-constrained rate distortion theory. In particular, the concept of set partitioning developed for TCM is applied to the partitioning of an expanded codebook into disjoint subsets, which are then used to label the branches of a state diagram. The excellent mean-squa.re error (MSE) performance and the low complexity of TCQ suggest its usefulness for compression of planetary images. 2. Description of the Compression Algorithm. A typical planetary picture contains SO0 x 800 samples pixels), at pre-processed with a 2-dimensional discrete cosine transform (DCT or ICT) on a 8 x 8 block of the picture, thereby producing 64 frequency domain coefficients which need to be properly quant ized. The 64 frequency-domain coefficients are then collected into a grey level resolution of 8 bit/pixel. These pixe s are first