Remote Sensing Image Target Detection Based On SSD Convolutional Neural Network

Remote sensing image target detection has been widely used in both military and civil fields and is one of the research hotspots at home and abroad.The traditional remote sensing image target detection algorithm has disadvantages such as weak feature extraction ability and poor generalization ability,while the convolutional neural network algorithm solves the disadvantages such as low adaptability of the traditional method and poor robustness of feature extraction,and achieves good results in many practical target detection problems.SSD(Single Shot MultiBox Detector)algorithm is a relatively popular convolutional neural network at present.Considering both detection accuracy and detection speed,SSD algorithm is selected in this paper for remote sensing image target detection.However,due to the different imaging methods of remote sensing images and natural images,remote sensing images have the characteristics of large image size and small target,so the detection accuracy of remote sensing image target detection directly using SSD is relatively low.Aiming at this problem,the SSD algorithm is improved in this paper.The main work is as follows:(1)In view of the problem of image information loss during input of SSD algorithm due to the large size of remote sensing image,a sliding window mechanism is proposed to avoid the operation of SSD to scale the input image,and the image edge information is used to remove the redundant images generated by the sliding window.Experimental results show that the image information integrity can be maintained by sliding window mechanism,and the detection accuracy of sliding window SSD algorithm is higher than that of the SSD algorithm before the improvement.(2)Aiming at the problem of low accuracy of SSD detection due to small targets in remote sensing images,a feature fusion method was proposed to improve the accuracy of SSD detection of small targets.In this method,the low-level feature map is used to replace the high-level feature map with lower contribution,and the extracted high-level feature map is deconvolved,the deconvolved feature map is fused with the low-level feature map,and the fused feature map is sent to the detector for detection.Experimental results show that the SSD algorithm with the added feature fusion has better detection effect on small targets in remote sensing images,and its detection accuracy is higher than that of the SSD algorithmbefore the improvement.(3)The input of the feature fusion SSD is the original image,and there is still a problem that the image information is lost due to the zoom operation.To solve this problem,the mechanism of sliding window is added to the SSD network model of feature fusion to form the SSD network model of sliding window + feature fusion.Experimental results show that the detection accuracy of feature fusion SSD algorithm with sliding window is higher than that of feature fusion SSD algorithm.