Research On Super-Resolution Algorithm Of Remote Sensing Images Based On Convolutional Neural Network

Remote sensing images usually contain a large number of ground object information,which has a very important application value in vegetation cover detection,land monitoring,disaster monitoring,urban economic level assessment,resource exploration,forest fire monitoring,military target attack and other aspects.The resolution of remote sensing image is an important standard to measure the information contained in remote sensing image,and also one of the important indicators to measure the quality of remote sensing image.However,due to the limitations of hardware equipment,it is costly and difficult to improve the resolution of remote sensing images.Therefore,it is an important topic in the field of remote sensing image processing to improve the resolution of remote sensing images by using image processing methods from the aspect of software.Based on the popular deep learning and convolutional neural network,this paper studies the super resolution method for remote sensing images.The specific research contents and innovations are as follows:(1)Collecting remote sensing images using unmanned aerial vehicle in Nanjing Lile Farm,and building LILE244 data set with 244 sets of high-resolution and low-resolution images as part of the research object.(2)Aiming at problems such as increased training difficulty and performance degradation after deep learning model is deepened,a super resolution reconstruction method of remote sensing image based on residual network is proposed.The model is composed of three parts: shallow feature extraction,deep feature extraction and up-sampling.Based on the residual block in Res Net,it uses Batch Normalization layer and subpixel convolution in the up-sampling part.This method constructs a convolutional neural network model that is both "wide" and "deep".In the performance test results of general image,satellite remote sensing image and UAV remote sensing image,the performance index exceeds the comparison method,and the visual effect is good.(3)In order to further improve the performance of the algorithm,a super resolution algorithm model of remote sensing image based on bi-direction attention mechanism is proposed.In order to further deepen the model and store more prior information,this paper constructs a nested residual structure by using long jump join and short jump join.This paper proposes a location attention module,which overcomes the defects of channel attention mechanism and weights the feature graphs in both vertical and horizontal directions to make the information of the feature graphs more efficient and thus increase the performance of the model.The model is trained and tested,and its super resolution reconstruction performance of general image and remote sensing image exceeds the above method,and achieves better image reconstruction effect.(4)In order to improve the efficiency of the algorithm,reduce the number of parameters and calculation amount of the model while maintaining the performance of the model as much as possible,a super resolution lightweight convolutional neural network model based on the selective channel processing mechanism is proposed.Through the channel selection matrix,the importance of each channel is learned in the training process,and the calculation of unimportant channels is ignored in the reasoning process,thus reducing the number of parameters and computation of the model.Differentiated attention module is introduced to further improve the performance of the model.The model is trained and tested.Under the premise that the number of parameters is less than 1M,the super resolution reconstruction performance of the universal image and remote sensing image exceeds the comparison method,and the super resolution reconstruction effect is better.