Study On Super Resolution Algorithm For Stereo Image

Because of its real sensory experience, stereo image and video become more and more popular, and is widely applied to medical, military and other fields in the meantime. However they have multiplied cost and complexity of data processing than single image or video. In order to compress the data without affect the quality of visual perception, mixed-resolution video encoding based on the theory of the human binocular suppression technology is widely applied to stereo image and video. However, some views would be downsampled to the low resolution images, and they would reduce the visual experience of users.In this dissertation, we apply the stereo super-resolution(SR) technology to the low resolution view image, and restore its high-resolution clear image. It mainly include:(1) The research background and development of super-resolution technology for stereo image and the reconstruction quality evaluation standards are expounded. We also introduce the state-of-the-art super-resolution technology for stereo image:DIBR technology and stereo image super-resolution technique based on combining the low frequence with the high frequence.(2) We provide a new super resolution algorithm for stereo image based on NonLocal(NL) similarity. The downsampling processing of low-resolution view in the mixed-resolution encoding was modeled as the degraded imaging model, and the adjacent views were considered as the reference images. A high resolution projected view was synthesized as initial estimation using the reference images and depth information. Then we introduce the NL similarity regularization term into the degraded model. At last we used the gradient descent algorithm to find the optimal solution. The results of experiment indicate that proposed algorithm has higher reconstruction quality.(3) We provide a new super resolution algorithm for stereo image based on sparse representation and NonLocal similarity. Firstly, we extract low frequency information and the corresponding estimation error of high frequency information, and use K-mean algorithms for the sample clustering. And we train a coupled dictionary for each cluster and use the obtained dictionary to reconstruct the estimation error of high frequency information. Than we combine it with the low frequency of original image and high frequency of projected image to obtain the initial image, and use the NL algorithm proposed in the third chapter to process the obtained image and get the reconstruction image. The experimental results show that the proposed algorithm has higher reconstruction quality.