Research On Stereo Image Super-resolution Algorithm Based On Deep Learning

Stereo image super-resolution technology aims to recover high-resolution stereo images with sharp details from low-resolution stereo images,which has wide applications in real scenarios.In recent years,deep learning based stereo image superresolution has made great progress,but there are two main problems in existing methods.First,due to the low quality of stereo image datasets,the spatial information extraction ability of existing models is limited.Second,when capturing the correspondence between stereo images,the existing methods only consider the effect of horizontal parallax and ignore the possible vertical parallax.This causes some pixels to fail to match the most similar pixels in the other view.To address these problems,this thesis proposes two stereo image super-resolution algorithms based on deep learning.First,this thesis proposes a stereo image super-resolution algorithm based on knowledge distillation.The single image super-resolution network trained on a more complete dataset is used as the teacher network and the stereo image super-resolution network is used as the student network.By transferring the knowledge of the teacher network,the student network can obtain more information and improve its spatial information extraction ability.Specifically,an adaptive residual feature aggregation network is proposed as the student network,which is composed of residual multi-scale aggregation module,parallax attention module and adaptive residual aggregation module.Moreover,a pixel-wise distillation strategy is proposed,which uses the output image of the teacher network as the soft target.The training process of the student network is supervised by calculating the pixel-wise differences between the student network output image and the teacher network output image.The experimental results show that the proposed knowledge distillation strategy significantly improves the performance of the student network,with a performance gain of about 0.2d B in terms of Peak Signal-to-Noise Ratio(PSNR).And the student network under this strategy has achieved the optimal effect on all test datasets.Second,this thesis proposes an enhanced back-projection stereo image superresolution network considering vertical parallax.Specifically,in order to capture vertical and horizontal parallax in stereo images,a relaxed parallax attention module is proposed.By calculating the similarity of pixels in a larger range,this module can alleviate the effect of vertical and horizontal parallax on stereo matching.Moreover,an enhanced back-projection block is proposed as the basic unit of the network.Through repeated up-sampling and down-sampling operations,this block can learn the interdependence between high-and low-resolution images and further improve the quality of image reconstruction.Experimental results show that the PSNR of the proposed algorithm is 0.14 d B higher than that of other stereo image super-resolution algorithms on the 4× super-resolution task,with faster convergence rate.