Research On Object Detection Algorithm Of Remote Sensing Image Based On Deep Learning

With the rapid development of aerospace technology and remote sensing technology,more and more specialized remote sensing satellites have been put into orbit,which means that the remote sensing image data available to us has increased exponentially.Remote sensing images usually contain a large amount of information,which can be widely used in environmental monitoring,urban planning,resource exploration and military target detection.However,the complexity of remote sensing image is becoming more and more complex,the range of resolution is becoming larger and larger,and the coverage of the image is becoming wider and wider.Moreover,the features of objects of the same type are often different in different remote sensing images,so it is of great significance to detect useful information quickly and accurately in remote sensing images.Different from natural scene images,remote sensing images have special characteristics.Remote sensing images usually contain a large number of small targets and dense scenes,as well as problems such as multi-scale targets and complex background.Therefore,more targeted algorithms need to be designed to improve the accuracy of remote sensing image target detection.In recent years,deep learning target detection algorithm has shown good performance in target detection tasks.Compared with traditional target detection algorithm with complex process,slow detection speed and poor robustness,deep learning algorithm has been significantly improved in all aspects.Among them,the method based on convolutional neural network shows high performance,so the research direction of this thesis is to improve the accuracy of convolutional neural network in remote sensing image target detection.In this thesis,the existing difficulties in remote sensing image target detection are analyzed,and an improved model method is proposed based on the target detection algorithm YOLOv5 to achieve higher detection accuracy.The main research contents include the following two aspects:(1)In remote sensing images,a remote sensing image detection method based on improved YOLOv5 is proposed to solve the problem of limited feature information and difficult detection due to the small target image coverage area.Firstly,adding the cooperative attention mechanism can effectively embed the target location information into the model channel attention network,and enhance the model’s cooperative location perception ability to detect the target information.Secondly,the context conversion module is introduced into the backbone network to better take into account the extraction of global information and context information,and improve the feature extraction ability for small targets.Finally,the frame position regression loss function is improved to improve the accuracy of remote sensing target location recognition and prediction box.Finally,a series of experiments were carried out on RSOD and UCAS-AOD data sets,which are mostly small targets,to verify the effectiveness and superiority of the improved algorithm.(2)In order to solve the problems of low detection accuracy caused by complex background and diverse target scales in remote sensing images,this thesis improved the YOLOv5 network structure.Deformable convolution of 7-inch convolution kernel is used to extract features of the target layer,which minimizes extraction of irrelevant features and enhances anti-interference performance of the model.Feature extraction and feature fusion of the large target layer are carried out by using the depth separable convolution kernel of 31-inch convolution kernel,which not only increases the effective receptive field but also reduces the amount of model calculation.By improving the sensitivity field and feature extraction ability of the large target layer of YOLOv5,the detection accuracy of medium scale targets and large scale targets in remote sensing images can be improved.Experiments were carried out on the remote sensing data set NWPU VHR-10,which has various categories and scales,and the improved algorithm has a remarkable enhancement effect for large and medium-sized targets.Experimental results show that the improved model can effectively improve the accuracy of multi-scale target detection in remote sensing images.