Research On Super-resolution For Stereoscopic Images

Super-resolution technology is designed to recover high-resolution images or videos from low-resolution images or videos,resulting in higher pixel density,richer detail,and sharper edges for the reconstructed images or videos.This technology has been widely used in many fields such as satellite remote sensing,medical imaging,video surveillance,and so on.Compared to single image,stereoscopic images contain two images,including the left view and the right view.The super-resolution of stereoscopic images not only needs to utilize the information of the left image and right image,but also needs to use the complementary information between two images to recover high resolution stereoscopic images according to the characteristics of the stereoscopic images.Therefore,the stereoscopic image super-resolution has important research significance.Based on the analysis of the key issues in stereoscopic image super-resolution,this paper focuses on stereoscopic image super-resolution based on deep learning to improve the super-resolution performance.This thesis proposes a stereoscopic image super-resolution method based on interactive module.The method consists of three components,including the spatial component,the interaction component,and the reconstruction component.Firstly,the spatial information of the left image and right image are extracted in the spatial component to realize the preliminary feature extraction.Then,through the interaction part,the information in the left view is used to assist in generating a high-resolution right view,while the information in the right view is used to assist the super-resolution of the left view.In particular,this chapter designs a stereoscopic image interaction module that selects information while retaining dual viewpoint information,thereby achieving full utilization of stereoscopic information in an interactive manner.Finally,through the reconstruction part,the final high resolution stereo image is obtained.Experimental results show that the proposed method achieves better reconstruction performance on the public data set.This thesis also implements a stereoscopic image super resolution method based on parallax attention.Firstly,the residual channel attention module is used as the basic network of single image feature extraction,and the channel attention mechanism is embedded in the module to adaptively adjust the channel weight.Then,the parallax attention module is used to find the matching pixel points in two images and promotes mutual learning of information between the left view and the right view.Through this component,the parallax offset can be effectively avoided.Finally,the final high resolution stereoscopic images are obtained by the upsampling module and the reconstruction module.Experimental results on the public dataset show that the implemented method can obtain good super-resolution reconstruction performance.