Candidate: Iker Vasquez – Computer Science Emphasis
Title: Sinusoidal Projection for 360 Degree Image Compression and Triangular Discrete Cosine Transform Impact in the JPEG Pipeline
Program: Doctor of Philosophy in Computing
Advisor: Dr. Steven Cutchin, Computer Science
Committee Members: Dr. Sole Pera, Computer Science, and Dr. Elena Sherman, Computer Science
Using the equirectangular projection is not efficient to store and transmit 360 degree due to their topography. When the captured pixels in the 3D space are projected onto the 2D plane using the equirectangular projection to form the 360 degree image, some of the pixels are stretched generating redundant data. These extra pixels in the image add extra memory requirements that have low impact in the final image quality. In this manuscript we use the sinusoidal projection with pixel rearrangement to compress 360 degree images. This approach removes the redundant pixels while keeping a high image quality in the spherical domain, showing better compression ratios than the equirectangular projection. In addition, we introduce the TDCT as a replacement of the DCT in the JPEG pipeline. The sinusoidal projection generates warped images towards the image corners and therefore, the frequencies within the image are not aligned with the square DCT. Instead of transforming squared image tiles, we generate and transform triangular tiles to analyze the compression impact. In order to use triangular tiles within the JPEG pipeline, we modify certain JPEG steps and present specific improvements to enhance the use of TDCT in the compression scheme. As a consequence, we got mixed results when compressing images with this technique: we perform worse over low noisy image but we achieve better compression ratios over high noisy images. This opens a new research line where can take advantage of triangular adaptability to improve the standard compression algorithms.