A wireless multimedia sensor network (WMSN) requires a video encoding system that should be energy-efficient because of its special characteristics: limited power capacity but long service life without maintenance. Besides, video is one of the most important information contents that a WMSN delivers. Motion estimation plays an important role in predictive coding, which is the key part of video encoding and requires large amount of computation. In order to reduce the computational complexity of motion estimation, the block-matching search algorithm is able to find the motion vectors using fewer search points. An algorithm which adds a predictive search technique to the enhanced modified orthogonal search (EMOS) has been proposed. It improves the efficiency of block-matching search algorithms, and has two self-adapting search patterns for large and small movements. The proposed algorithm requires less search points to work out the movement of blocks and provides acceptable image quality. This algorithm was also tested on field programmable gate array (FPGA) and Arduino platforms.
Moreover, a back propagation neural network model is introduced for predictive block-matching. The proposed back propagation neural network has very simple structure with only 5 inputs, 5 hidden neurons and 1 output architecture. Because of its simplicity, it requires very little computational power which is negligible compared with existing computation complexity. The test results show the prediction accuracy in 10 - 30\% higher then the competing algorithms with a peak signal-to-noise ratro (PSNR) improvement up to 0.3 dB. The above advantages make it a feasible replacement of the current solution.
With the information technology developing dramatically, there is reason to believe that the next generation video encoding standard HEVC will soon be able to run on very cheap platforms. Therefore, a prospective study on HEVC inter prediction acceleration was also carried out. We extracted specific image features that represent prediction unit texture, incorporated a machine learning technique, namely random forest, in HEVC intra prediction mode selection, to improve the performance of inter coding of HEVC. Benchmarking with other existing algorithms, our method extracts very specific features of image texture changes in terms of angle. Therefore the proposed method can achieve very high prediction accuracy. Having similar reduction in complexity, the proposed algorithm will be demonstrated to have a higher video quality compared with similar algorithms.
|Date of Award||28 May 2020|
- Univerisity of Nottingham
|Supervisor||Sherif Welsen Shaker (Supervisor) & Sui-Yeung Cho (Supervisor)|
- wireless multimedia sensor network
- video processing