Contributions to data compression research for various media formats using particle swarm optimization and neural networks
Date of Award
Doctor of Philosophy (PhD)
Electrical and Computer Engineering
W. David Pan
B. Earl Wells
Sivaguru S. Ravindran
Data compression (Computer science), Image compression--Standards, Video compression--Standards
Advancement in high resolution devices used to capture images and videos has put strain on data storage and transmission. For their practical usage, efficient data compression techniques becomes essential. In this context, we conducted a study on compression of two media formats, including ROI maps of hyperspectral images, and 3D point clouds. For lossless compression of ROI maps, we proposed novel schemes that can achieve high compression ratios by partitioning the intricate regions of the image into smaller blocks. Our analysis of a linear predictive model applied on the original images showed that the resulting residual images tend to have lower entropy. We then proposed to use the DPSO algorithm that searched for the best combination of scan directions to achieve better compression. Simulation results on various data sets showed that the proposed scheme outperformed JBIG2, the international standard for binary image compression. Our work also advanced the research on lossy attribute compression of 3D point clouds in two aspects. First, we proposed an adaptive scheme based on 3D Sobel filters that can switch between the RAHT and Dyadic RAHT to achieve early termination. We showed that the proposed method was able to achieve noticeable compression gains over the all-dyadic approach. Second, we further improved the adaptive method by training a neural networks so that the switching would no longer depend on a threshold. Compared to the GPCC standardized method that uses only Dyadic transform throughout the point cloud, the proposed scheme can achieve cumulative BR-rate gains on various publicly available datasets.
Hooda, Reetu, "Contributions to data compression research for various media formats using particle swarm optimization and neural networks" (2023). Dissertations. 274.