Research On Small Target Detection Methods In Optical Remote Sensing Images

The detection of targets in optical remote sensing images plays an important role in both civil and military Settings.With the continued progress of digital imaging technology and computer vision technology,lens detection methods for optical remote sensing imaging have gradually become a popular topic.However,unlike natural images,there are some unique characteristics and challenges for detecting targets in optical remote sensing images.Firstly,optical remote sensing images are densely arranged,which poses difficulties for target detection and recognition.First,optical remote sensing images are densely arranged,which poses difficulties for target detection and recognition.Second,the size of targets in optical remote sensing images varies considerably,with some targets being very small while others may be very large,requiring greater robustness of the algorithm.Based on this extensive research background,the main research of this paper is:(1)This method is a target detection method based on the YOLOv4 attention mechanism,which takes into account the characteristics of optical remote sensing images with many small-scale targets and dense arrangements.The innovation of the method is the introduction of the CBAM attention mechanism(Convolutional Block Attention Module,CBAM),which is applied to the residual module of the backbone network to give more information about the characteristics of smallscale targets and remove irrelevant background information,thus improving the generalization ability of the model.The method also optimizes the NMS network to improve the localization accuracy of detected targets by finding the optimal bounding boxes and removing unnecessary redundant boxes.Finally,the method also improves the sensing head part of the network model by adding a small target oriented sensing head to improve the detection efficiency of small targets.Experimental results show that the method achieves 78.1% mAP on the DIOR dataset,which has higher accuracy and better robustness compared to typical target detection methods.(2)The method is based on the YOLOv4 model with light improvements to the support network.In particular,the mobilenetv3 light network is used instead of the original signature extraction network,reducing the number of parameters and complexity of the network.In addition,to further improve the model’s detection speed,the method also USES depth separable convolution to replace the traditional neck network convolution of the remote sensing optical lens detection model.Experimental results show that the method achieves a detection speed of 58 FPS with guaranteed detection accuracy,which is faster and more efficient compared to the traditional method of target detection by optical remote sensing.Thus,this method can significantly improve the efficiency and accuracy of detection of optical remote sensing objectives in practical applications and has broad application prospects.