Researches On Multi-Images Super-Resolution Based On Deep Learning

As a low-level computer vision task,image super-resolution can provide assistance and enhancement for higher-level computer vision tasks.However,single-image super-resolution technology can only rely on the information in the image to reconstruct high-resolution image,and the ability to reconstruct image details is limited.Therefore,multi-images super-resolution technology has received more and more attention.Among them,the use of stereo images and image sequences for image super-resolution has a wide range of application values in industrial production and daily life.This paper proposes state-of-the-art methods for stereo image and image sequence super-resolution,respectively,to advance the field.The traditional stereo image super-resolution method has many limitations,such as: the convolutional neural network cannot fully extract the joint representation of the local and global features of the image,and the algorithm for pixel-level alignment of stereo images using parallax disparity is not robust.Therefore,this paper studies and proposes a Transformer method based on feature matching and fusion to solve the above problems.In addition,traditional image sequence super-resolution methods cannot effectively integrate multi-source information with differences.To this end,we adopt an early fusion method to integrate information from different information sources,and propose a model based on Siamese recurrent neural network,which achieves a very significant performance improvement by solving the above problems.Specifically,the main research contents and contributions of this paper are as follows:1.A Transformer method based on feature matching and fusion is proposed for stereo image super-resolution.By using a self-attention-based Transformer as a feature extraction backbone network,a joint representation of rich and robust local and global information within an image is sufficiently extracted.In addition,by combining correlation coding,hard attention and soft attention techniques,a feature matching fusion module is proposed,which can achieve sufficient information interaction between left and right views while avoiding complex pixel-level alignment operations.Based on the above design,the method proposed in this paper achieves state-of-the-art performance on four common public test sets.2.A Siamese recurrent neural network model is proposed for image sequence super-resolution.Firstly,by adding neighboring low-resolution images and the previous super-resolution results as prior information to the network input,the reconstruction results of the current moment contain more information of image types.Then,the Siamese network is used to extract deep features,and a feature fusion module based on spatial and channel attention is used to integrate multi-source information,which solves the problems of spatial misalignment,distribution difference,and domain shift that exist in multi-source information.Based on the above design,the method proposed in this paper achieves competitive performance with representative research results on two common public test sets.