A novel block-based motion estimation algorithm and architectures for inter-frame video coding

Current video compression standards such as MPEG-1/2 and H.261/3 make it possible to develop low cost, high performance, real time visual communication systems to support emerging applications such as multimedia, video-on-demand (VoD), and high definition television (HDTV). These applications demand much higher video compression performance than previously available, and it is widely accepted that in order to implement these new applications in real time, special purpose VLSI circuitry is the only answer. Common to the current video compression standards is the operation of motion estimation which exploits temporal redundancy among successive frames of a video sequence to reduce the amount of data needed to be transmitted. This process is the most time consuming procedure in these video compression standards. In this dissertation, we propose modular systolic VLSI array structures to implement full search block matching algorithm and logarithmic search block matching algorithm. A key step in this work is to exploit the patterns of overlapped search area in the reference frame so as to reduce as much as possible the needs to reload or redistribute the same reference data from an off-chip buffer over and over. As a consequence, our architectures offer more efficient processor utilization while achieving very high throughput rate and minimum number of input ports. On the software side, we introduce a new matching criterion and a simple image segmentation technique based on variable block size to reduce computational and hardware complexity at the cost of moderate performance degradation in terms of peak signal to noise ratio or total compression ratio. A key step in this work is to simplify the matching criterion for the block-based motion estimation algorithm to reduce hardware complexity and number of input ports. Furthermore, to alleviate the problem of massive amounts of computations involved in the conventional motion estimation algorithms, we propose a motion estimation algorithm based on variable block size in this thesis. This proposed algorithm estimates only a subset of the eligible target blocks in the current frame, and minimizes the prediction error of the region which involves complex motions.