Multi-class Object Detection For Remote Sensing Images

In recent years,with the development of image processing technology,remote sensing object detection technology has been widely used in military and civilian fields such as military reconnaissance,missile early warning,and urban planning.Scholars at home and abroad have done a lot of research works on remote sensing object detection,the most popular one is the remote sensing object detection algorithm based on deep convolutional neural networks.However,with the large-scale increase of remote sensing image datasets,remote sensing objects gradually show characteristics of multiple classes,multiple angles,and multiple scales.But most existing remote sensing object detection networks only detect specific objects,which have low versatile and weak capability of feature extraction,unable to accurately detect multiple types of multi-angle objects,and small-scale objects are difficult to detect.Therefore,the research on multi-class object detection of remote sensing image technology has important practical significance and wide application prospects.In the existing deep convolutional neural networks,because the Region Proposal Network(RPN)can generate object candidate boxes and reduce the amount of network parameter,it is widely used in remote sensing object detection.However,current existing region proposal network usually generate horizontal candidate boxes,which will cause detection error and reduce detection accuracy when extracting multi-angle object features.At the same time,because the range of feature extraction in the convolutional layer can be controlled by the receptive fields with different sizes,so multi-scale object detection often use different sizes receptive fields to process feature extraction.However,when the receptive fields of various sizes act on all sizes of objects,the calculation will be huge and not targeted.This article studies the above two issues,and the specific research contents are as follows:(1)Aiming at the problem of high error rate of multi-angle remote sensing objects detection,the architecture of multi-angle object candidate boxes is researched,and a Rotation-guided Object Proposal Network(ROPN)is proposed.An anchor frame that matches the shape of the object is generated by predicting the center point,angle,and size information of the object,this method reduces the generation of a large number of redundant anchor frames,and a feature map correction module is added to adapt the feature map of the convolution layer to the shape of the anchor frame,thereby generating a more accurate object candidate boxes.At the same time,the algorithm also introduces a multi-angle pooling layer,which can process feature extraction on multi-angle candidate regions of different scales,meanwhile further improving the detection accuracy of multi-angle objects.(2)Aiming at the difficulty of detecting small objects in multi-scale objects of remote sensing images,we researched the influence of receptive field on the detection accuracy of objects at different scales,thus a multi-scale object detection algorithm based on parallel receptive field branch structure is proposed.The structure first sets up parallel small convolution layers to increase the network width.After each convolution layer whose convolution kernel size is greater than 1,three parallel dilated convolutions with different dilation rates are added to generate receptive field regions with large,medium and small sizes.By setting area thresholds for all objects,they are divided into receptive field regions of corresponding size to process feature extraction.Receptive field branches adopt a weightsharing strategy,which improves the accuracy of object feature extraction while reducing the amount of parameter calculation,further improving network computational efficiency.At the same time,the output features of the original network layer assist the output features of the parallel receptive field branch structure through feature fusion to further improve the accuracy of multi-scale feature extraction.Based on the research of multi-angle and multi-scale remote sensing object detection,this paper proposes a multi-class,multi-angle,multi-scale remote sensing object detection algorithm by combining the Rotation-guided Object Proposal Network(ROPN)and the parallel receptive field branch structure.First,the algorithm in this paper is compared with the current advanced remote sensing object detection algorithms.Then the validity of each structure of the algorithm is verified through multiple sets of experiments,and the generalization ability of the algorithm is tested through different datasets.The effectiveness and robustness of the multi-class object detection algorithm for remote sensing images are presented.Finally,based on the research results of multi-class object detection for remote sensing images,a software based on Web platform is designed and implemented,which achieves efficient detection of multi-class,multi-angle,multi-scale remote sensing objects.