Study Of Key Techniques In Compression Of Ultra-high-definition Images With Constrained Resource

In the past decade, the acquisition, transmission and storage of digital images have become pervasive in our daily life. Digital technology allows digital images to be regenerated, processed, archived and transmitted easily. Most significantly, a diverse range of service booming over Internet fit in with digital images readily. Despite their advantages,there is one problem with digital images – huge volumes of data involve when they are represented in uncompressed data form, especially for those ultra-high-definition image data produced by highly sensitive image sensors. Image compression is an indispensable tool for efficient representation, storage and delivery of digital images.For a very long time, the main goal of image compression has been to improve the compression gain, while more computational/memory requirements were taken for granted on account of the rapid development of the VLSI technology. Nowadays, however, the attributes to image compression techniques have changed. Mobile and wireless devices, e.g., smartphones or small sensors network, evolved and became widely accepted, require algorithms with highly constrained computational power and memory budgets, while the resolution/dimension of the images to be handled increases steadily.For these devices there exist requirements aside from compression gain, including low power consumption, high scalability, low complexity, and low memory. For example, although the wavelet-based image codecs have been shown outperform their DCT counterparts for a very long time, the implementation complexities have limited their widespread use in consumer electronics, especially in hand-held and battery-operated devices.Typical image compression systems include two building blocks, i.e., decorrelating transform and codec, which are addressed sequentially in the thesis. The second chapter of this thesis concerns the design and implementation of such discrete decorrelating transforms(i.e., the block transform, the lapped transform and the discrete wavelet transform), put into the scenario of ultra-high-definition images. To cope with the memory bottleneck posed by the massive volume of image pixels, low-memory and low-complexity transform implementation techniques are proposed in the third chapter. Considering that low end-to-end delay is critical to real-time image/vido communication(e.g. real-time image compression by spy satellites), the fourth chapter addresses the minimum end-toend delay of generic two-channel linear-phase perfect reconstruction filter banks. In the fifth chapter, state-of-the-art image codecs are surveyed, based on which an image codec adapted to ultra-high-dimension images is proposed to provide an integral image compression system which could process ultra-high-definition images readily.The contributions of the thesis are fivefold, as detailed in the following.Based on the decomposition of flipping structure into two phases of even and odd time indices, a novel low-memory and low-complexity implementation for the discrete wavelet transform called the DFS(Decomposed Flipping Structure) is proposed. Compared to the traditional implementation of lifting scheme, the DFS consumes less memory and has lower computational complexity. For the CDF 9/7 wavelet filter bank widely adopted by wavelet-based image/video coding schemes, the conventional real-time lifting scheme requires six memory cells(IO units included), or five memory cells with an 100%increase of multiplication number. In contrast, the proposed DFS provides an in-place implementation with five memory cells, maintaining the same computational complexity. The experimental results show that the DFS has a speedup of 44% compared to the conventional real-time implementation of lifting scheme.Based on the optimization of causal implementation for lifting schemes and interleaving modes of subband coefficients, this thesis presents an enhanced low-memory implementation of discrete wavelet transform called the Opt-LBWT(Optimized Line-Based Wavelet Transform). In contrast with the conventional global implementation of discrete wavelet transform, the Opt-LBWT has the advantage that its memory budget is independent of the image height. When the difference between the filters’ lengths is greater than two, the Opt-LBWT has lower memory requirement and system latency than the linebased wavelet transform(LBWT). Taking as an example, when two-dimensional 5-level decomposition with the MPEG-4 Default 9/3 filter bank is adopted, the overall memory reduction is 22.7% compared with the LBWT.An on-the-fly and memory-scalable implementation of hierarchical lapped transforms with minimum memory budget is proposed. Both the forward and inverse transforms have the advantages of low memory requirement unrelated to the length of the input signal. Experimental results suggest that the proposed algorithm outperforms the existing algorithms with respect to computational efficiency and memory requirement.Discrete multiresolution transforms based on two-channel linear-phase FIR filter banks and their variants are used ubiquitously for coding natural images/videos. In realtime video communication(e.g., the video conferencing/surveillance), constraining the end-to-end delay incurred by temporal multiresolution transforms is critical for timely response. In this thesis, the end-to-end delay of the analysis/synthesis banks for generic two-channel linear-phase FIR filter banks is minimized and proved by analytically formulating the temporal dependency between the reconstructed frames and the input frames.As a result, the closed-form minimal end-to-end delay is attained and proved. In addition,a DWT implementation with minimum momery budget and minimum end-to-end delay called the Mini-Delay DWT is proposed.An integral image compression system that could process ultra-high definition/dimension images with limited memory and computational resources is presented to show the efficiency of the above-mentioned techniques. Competitive performances are observed with respect to the state-of-the-art image codecs which are the most memory and computation intensive in the literature.