Toward Assessing and Improving the Quality of Stereo Images

Imaging systems have incorporated numerous technological innovations such as 3-D television and handheld devices. Despite these advances, these techniques still require the human eyes to refocus until the sense of depth perception is achieved by the observer. The more time this takes, the more eye muscles become fatigued and the brain tires from confusion. However, the exact intricacies involved are far more complex. To alleviate these problems, we introduce a learning framework that aims to improve the quality of stereo images. Instead of attempting to cover all factors that affect the quality of stereo images, such as image resolution, monitor response, viewing glass response, viewing conditions, viewer differences, and compression artifacts, we first introduce a set of universally relevant geometric stereo features for anaglyph image analysis based on feature point correspondence across color channels. We then build a regression model that effectively captures the relationship between the stereo features and the quality of stereo images and show that the model performs on par with the average human judge in our study. Finally, we demonstrate the value of the proposed quality model in two proposed applications where it is used to help enhance the quality of stereo images and also to extract stereo key frames from a captured 2-D video.