Master Thesis: A Deep CNN for the Quality Enhancement of Stereo Images

Speaker: Qiuhai Guo

Room: 0406

Stereo images can provide users with immersive watching and gaming experiences. However, they also bring higher transmission and storage costs compared to single-view images. Using different compression ratios on the two views is a potential solution to this problem, but the compressed images reduce the stereo image quality, and this process is irreversible. Since stereo images contain a lot of redundant information between two views, it is possible to enhance the low-quality (LQ) images using the high-quality (HQ) images. To take advantage of the redundant information of a stereo image, it is necessary to estimate the spatial translation between the left and right views.

In this work, a deep CNN based approach is proposed to solve these problems. With pre-processing and designed convolutional neural network, we can effectly enhance the LQ view with the guidance of HQ view images.

Supervisor(s): Dr. Zhi Jin, Kai Cui