Contributions To Image Compression With Low-Memory Requirement

As an effective measure to alleviate the pressure of storage and communication chan-nel, image compression has been an active research field in recent years. With the ad-vancements of high-sensitivity sensors, large-size images have become ubiquitous. There is a urgent need to compress these images with limited memory. The market of con-sumer electronics (e.g. digital cameras and cell phones) has grown rapidly. The on-chip memory of those devices is severely limited and is a critical element in hardware cost. Implementing image compression on these memory-constrained environments is a urgent research problem. Due to their superior performance of energy compacting and redun-dancy removal, Discrete Wavelet Transforms (DWTs), Generic Hierarchical Transforms (GHTs), and Generic Tree-Structured Filter Banks (GTSFBs) are amongst the best dis-crete transforms and have been utilized widely in image compression systems. Taking into account of the key functions of discrete transforms in image compression, in this thesis we mainly address the low-memory or memory-scalable implementations of these three classes of discrete transforms. The novelties of this thesis could be summarized as follows.Based on the Lifting Scheme and the compact support property of popular DWT filter banks, a low-memory implementation of DWT called the stripe-based wavelet transform (SBWT) is proposed, which produces the same subband coefficients as the global imple-mentation of DWT. The memory budget of SBWT depends only on the DWT filter banks adopted, the number of the transform levels and the width of the image. Additionally, the production of subband coefficients lends itself to PCT (Parent-Children Tree)-based image coders in that no intermediate buffering is needed between SBWT and the coders. Taking into account that different PCTs correspond to different image regions, the PCT coder utilizing the SBWT can achieve random access of images easily and has high error-resilient property. When coupled with PCT-based image codec's and 5-level decomposition with the CDF 9/7 filter bank, the overall memory reductions are 18.4% and 17.9% compared with the line-based DWT and the memory-constrained WT, respectively.Based on the compact support property of popular multi-channel filter banks and the FIFO (First-In First-Out) cache technology, a unified approach for low-memory and on-the-fly implementation of GHTs, called the stripe-based hierarchical transform (SBHT), is proposed. The SBHT has four advantages:1) It operates at the same time while the data are being acquired and that all the subband coefficients should have been generated by the time the data acquisition ends.2) It generates the same subband coefficients as the traditional global implementation of GHTs with a low memory budget that only dependent on the GHT and the image width.3) It lends itself to PCT-based codecs in the sense that the basic data unit generated by the SBHT is row-based collections of PCTs, the most widely used data entity in subband coding. No intermediate buffering is needed between the SBHT and the codec that consumes PCTs.4) The memory budget needed by the SBHT is the minimal one for the PCT production. By analyzing the dataflow of different levels of GHT, the relation between the samples and the subband coefficients is also attained, by which the correctness of the SBHT is strictly proved.By reformulating the multi-channel filterbank as single-input single-output linear pe-riodically time-varying system and establishing a one-to-one mapping between subband co-efficients and time indices, a memory-scalable implementation of GTSFBs is presented that produces exactly the same subband coefficients as the global implementation of GTSFBs. Both the forward and inverse transforms of the proposed approach have low-memory re-quirements that are unrelated to the height of images and all the samples (or subband coefficients) only need to go through the transform engine once, no data rewinding needed. Additionally, the majority of the memory budget can be distributed freely between the forward and inverse transforms, which is highly desirable for the applications where the memory asymmetry exists between the analysis/synthesis systems.