Research On Remote Sensing Target Detection Algorithm Based On Attention And Feature Fusion

With the continuous progress of computer computing power,deep learning technology has been greatly developed,and computer vision technology is more applied to all aspects of life,such as object detection,face recognition,intelligent driving and so on.The rapid development of electronic equipment also makes the resolution of satellite remote sensing images higher and wider.Therefore,using deep learning method to process satellite remote sensing images can grasp the information in the images more accurately,and object detection for remote sensing images has become the focus of research.Compared with traditional algorithms,depth based object detection technology can label features more simply,and can rely on powerful data sets to improve training samples to achieve more efficient object recognition and more accurate position labeling.Therefore,depth learning based object detection method is of great significance for remote sensing image recognition.Aiming at the characteristics and detection difficulties of remote sensing images,in order to make the network better extract the features of objects in remote sensing images,firstly,aiming at the problems of small receptive field and insufficient feature extraction in the classical networks such as SSD and yolov3,a network based on channel and spatial attention is proposed in this thesis.Firstly,the residual blocks used for feature extraction are connected by cross stage partial network,which can enhance the learning ability of convolutional neural network.Then the combined channel and spatial attention module are connected to improve the feature extraction ability of the network.Finally,the extracted features are connected to the pooled pyramid network to integrate local features and global features to increase the receptive field of the network.After using the channel and spatial attention module,the network can better understand the image and pay more attention to important levels in feature extraction.The detection effect is improved,and the m AP is 67.7% on the Dior of remote sensing image data set.On this basis,in order to enhance the detection effect of multi-scale targets in complex background,this thesis applies the channel and spatial attention module to yolov5 network,and proposes hybrid data enhancement combined with mosaic and mixup data enhancement to expand the number of samples,increase the diversity of samples and enrich the background of samples.A multi feature pyramid network is proposed.The multi-scale fusion method is used to fuse the deep feature map with a large amount of semantic information and the shallow feature map with rich location information for many times,so as to obtain higher small target detection ability.Combined with the disadvantage that channel fusion treats input equally,a weighted channel fusion is proposed.The channel weight is added to the last fusion of multi feature pyramid network,so that the multi-scale feature map pays more attention to the important channels.After optimizing the loss function of object category classification and boundary box position regression,a multi-scale object detection algorithm based on weight feature fusion is obtained.Through experimental analysis and comparison,it can be concluded that the multi-scale target detection algorithm based on weight feature fusion proposed in this thesis has a m AP of 73.8% on the remote sensing image data set Dior,and has a more significant effect on small targets and multi-scale targets in complex background.