Anti-occlusion Remote Sensing Image Airplane Object Detection Method Based On YOLOv4

Remote sensing image object detection is one of the important directions of remote sensing image application.With the significant increase in the resolution of remote sensing images,the detection of a large number of artificial movable targets becomes possible.As the representative artificial mobile target,airplane targets can be quickly detected and recognized based on remote sensing images,which has important application value in civil and military fields.Because the airplane is movable and the remote sensing images are limited by the acquisition conditions,the airplane targets on the remote sensing images are easily occluded by hangars,clouds and other parts,which can easily cause false detections or missed detections.Traditional object detection methods have made important breakthroughs in image processing,but it is difficult to meet the rapid detection of a large number of targets.In view of the excellent performance of deep learning methods in natural image object detection tasks in recent years,this paper will focus on the research of anti-occlusion object detection methods based on deep learning.The occlusion mentioned in this paper is all partial occlusion.In order to achieve accurate and rapid detection of occluded airplane targets,the representative method YOLOv4 is selected from the deep learning object detection methods for research and optimization.In order to enhance the robustness of the model,only non-occluded airplane target samples are used to train the model.The main research contents of this paper are as follows:(1)There are some problems in the remote sensing image,such as blurry contour and missing features of the occluded airplane targets.This paper studies the visible part of the occluded airplane targets,and proposes YOLOv4-ASFF method based on adaptive spatial feature fusion.In order to enhance the feature expression ability of the network,the PANet feature fusion network in YOLOv4 is replaced by the FPN network.On the basis of the image features fused by FPN,the ASFF structure is introduced for feature adaptive fusion.Convolution operation is carried out on the feature maps of each layer to obtain the corresponding weight coefficients.The feature maps of each layer are multiplied with the corresponding weight before fusion,so that the model can highlight useful features and effectively suppress useless features.The experimental results show that the improved method has a good ability of anti-occlusion object detection,but the generalization ability for non-occluded airplane targets is insufficient.(2)YOLOv4-ASFF method improves the anti-occlusion object detection performance of the model,but the occluded airplane targets and the non-occluded airplane targets have different requirements for shape information of low-level features and semantic information of high-level features,so feature fusion is difficult to improve the detection accuracy in both cases.Considering that the target and the surrounding objects always appear together,this paper studies the global and local features of the images,and proposes a YOLOv4-RFEM method based on receptive field enhancement mechanism.In order to increase the multi-scale receptive field,the receptive field enhancement mechanism is constructed by combining SPP and Dense ASPP.The dilated rate is designed by HDC strategy to reduce the gridding issue of atrous convolution.In feature fusion,the top-level feature maps generated by feature extraction network are replaced by multi-scale receptive field feature maps,and then fused with low-level feature maps.The experimental results show that the improved method can improve the detection ability of the model for both occluded and non-occluded airplane targets.This thesis has 38 figures,7 tables and 106 references.