Research And Implementation Of Monocular And Binocular Super Resolution Algorithm Based On Residual Learning And Attention Mechanism

Super resolution algorithm aims at recovering high-resolution images from low-resolution images.Usually,equipments in real life like medical equipment,remote sensing satellite and monitoring equipment have limitaions in accurary,which could not satisfy people's demand for resolution.In order to resolve the problem,one solution is to obtain high-resolution images directly from devices.However,it is not only costly,but also complex.Therefore,it is necessary to generate high-resolution images through algorithm.The super-resolution algorithm needs to solve the following problems.Firstly,how to solve the problem of various kinds of pictures and different sizes of objects.Secondly,how to distinguish the texture area from the smooth area,how to locate the high-frequency details,and how to solve the problems of missing image details,blurred image,and repairing small texture.Finally,how to solve the problem of various disparity and the redundant information between left and right images.To address these problems,the main research innovations of this paper are as follows.(1)In order to solve the problems of various kinds of pictures and different sizes of objects,a monocular super-resolution algorithm based on multi branch hierarchical structure is proposed.The method inserts dilated convolution with different dilated ratio on each branch to obtain multi-scale feature representation.In this way,the generalization ability of the model is enhanced and the requirements for various scenes and sizes are meeted.In addition,the residual learning mechanism is utilized to convey the low-frequency information to the reconstruction module,which makes the network focus on learning the high-frequency information to get better performance.Experiments on multiple datasets show that the algorithm can learn the mapping relationships of multi-scale and various images and generate high-quality images.(2)Aiming at solving the problems of missing image details,blurred image and small texture repairment,a monocular super-resolution algorithm based on attention mechanism is proposed.The algorithm compresses the information of each channel and learns the relationship among channels.This channel attention mechanism increases the difference between channels,so that different channels have different characteristic information.In addition,the spatial attention mechanism is included,which can distinguish the texture area from the smooth area and locate the high-frequency details.Experimental results show that the algorithm generates clearer images.(3)Aiming at solving the problem of various disparity and the redundant information between left and right images,a binocular-super-resolution algorithm based on parallax and attention mechanism is proposed.The algorithm learns the parallax information from the binocular images and warps the right image to obtain the new corresponding left-view representation through the parallax information.The new left-view representation contains the information missing from the left image.Similarly,the left image can also be warpped by the parallax information to generate the new right-view representation.The experimental results show that this method can effectively utilize the complementary information of binocular image.The experimental results show that the monocular and binocular super-resolution algorithm based on residual learning and attention mechanism has reached the leading technical level.The generated high-resolution images are more accurate and can be widely used in various computer image tasks.