A Fast and Low Complexity Image Codec based on Backward Coding of Wavelet Trees Jiangling Guo, Sunanda Mitra, Brian Nutter, Tanja Karp Computer Vision and Image Analysis Lab Dept. of Electrical and Computer Engineering Texas Tech University Sunanda.Mitra@ttu.edu Abstract A new approach of backward coding of wavelet trees (BCWT) is presented. Contrary to the common “forward” coding of wavelet trees from the highest level (lowest resolution), the new approach starts coding from the lowest level and goes backward by building a map of maximum quantization levels of descendants. BCWT eliminates several major bottlenecks of existing wavelet-tree-based codecs, namely tree-scanning, bitplane coding and dynamic lists management. Compared to SPIHT, BCWT encodes and decodes up to eight times faster without sacrificing PSNR. At the same time, BCWT provides desirable features such as low complexity, low memory usage, and resolution scalability. 1. Introduction The embedded zero-tree wavelet (EZW) [1] image coding algorithm, developed by Shapiro, and SPIHT [2], developed by Said and Pearlman, form the benchmarks for new, high performance, wavelet tree-based coding techniques. Despite the superiority of the above coding algorithms, recent literature [3] addresses the lack of certain desirable features that could yield further improvements. Desirable features that are relatively difficult to incorporate within the traditional wavelet-tree-based codecs are resolution scalability, low memory usage, and computational efficiency in encoding and decoding. Although, during the last decade, a number of wavelet-tree-based codecs [4-9] have emerged, most of them can only enhance one of these desired features, quite often, at the expense of other features. Among them, LTW (Lower-Tree Wavelet) [6] is promising, as it possesses all three features at the same time by abandoning the popular bitplane coding and replacing it with a two-pass coding. However, LTW’s output stream is not rate-embedded and heavily depends on arithmetic coding, which makes it impossible to adapt for rate-embedding applications and difficult to parallelize without PSNR trade-off. In this paper, we introduce a new codec, Backward Coding of Wavelet Trees (BCWT) [10]. This codec not only possesses resolution scalability but also has significantly faster coding speed, lower memory usage, and lower complexity than all other wavelet-tree-based codecs we have studied. The encoder can also be easily adjusted to provide rate-scalability and other types of progression. Most importantly, these features do not come at the sacrifice of PSNR performance, because the BCWT coding principle is essentially identical to SPIHT. In the next section, the basic concepts as well as the algorithmic details of BCWT including a numerical example to illustrate the algorithm are presented. In section III, results of coding speed, PNSR of decoded images, and memory usage for BCWT, SPIHT, LTW, and JPEG2000 are compared for the “Bike” image. In section IV, future work and applications of BCWT are discussed. Proceedings of the Data Compression Conference (DCC’06) 0-7695-2545-8 /06 $20.00 © 2006 IEEE