| With the development of deep learning,great progress has been made in stereo image super-resolution reconstruction techniques.Different from a single image,a stereo image contains two views of the left eye and the right eye,and the super-resolution reconstruction process not only needs to consider the information in the monocular view,but also needs to consider the complementary information between the left and right views.Therefore,how to effectively utilize the information of these two situations has become a research hotspot in the field of stereo image super-resolution.After in-depth analysis of the characteristics of stereo images,this paper proposes a stereo image super-resolution reconstruction algorithm based on feature fusion and attention mechanism,which realizes the efficient use of intra-view and interview information.The main research contents are as follows:Aiming at the problem that the existing stereo image super-resolution algorithms cannot effectively utilize the in-view information,this paper proposes a stereo image super-resolution reconstruction algorithm with multi-level fusion attention network.First,the algorithm utilizes residual-intensive atrous convolution modules to obtain multi-scale information,and continuously fuses high-and low-frequency information.Then,the disparity attention mechanism is combined with the channel spatial attention mechanism to improve the accuracy of feature alignment by fully mining the channel and spatial information of features,which not only improves the distinguishing expression ability of features,but also effectively solves the occlusion problem.Next,the residual dense network and attention mechanism are used to mine the intra-layer and inter-layer information of features of different depths,which further enriches the high-resolution details of the features.Finally,upsampling and global residual connection are performed to reconstruct a high-resolution left eye image with rich high-frequency and lowfrequency information and better stereo effect.Experiments on three benchmark test sets show that the algorithm achieves very good performance,can reconstruct the real details of the image and the model parameters are low.The PSNR and SSIM results obtained on the KITTI2015 dataset are respectively better than those of the same algorithm.Improved by 0.18 d B and 0.003.In order to make full use of the information between left and right views and further improve the performance of stereo image super-resolution algorithms,this paper improves on the first work and proposes a multi-scale feature fusion and alignment network.First,the network upscales the input low-resolution left and right images to the target size through deconvolution.Then,a multi-scale fusion extraction structure is designed to process these enlarged images,which can further capture the global context information of features while extracting multi-scale features.Then,the network performs feature alignment at multiple scales to achieve more refined image registration by combining the deformable convolution alignment structure and the parallax attention mechanism,which makes good use of the inter-view information.Afterwards,the registered results are modeled in multiple dimensions,enabling deep mining of channel,spatial,and inter-layer information of features.Finally,the multidimensional modeling results and the low-resolution up-sampled left-eye image are globally residual connected to obtain the super-resolution left-eye image of the network.Experiments show that the PSNR and SSIM results obtained by this network are greatly improved compared with the first work,and the reconstructed images are rich in details and have a strong threedimensional sense. |
