| With the vigorous development of remote sensing technology,the acquisition of high-quality optical remote sensing imagery has become easier,and object detection based on remote sensing imagery has also been widely used in the fields of national defense,military and civil economy.In recent years,object detection based on deep learning has dominated the field of natural image object detection due to excellent feature extraction and representation capabilities and has achieved remarkable detection results.However,the objects in remote sensing imagery have unique characteristics,which are very different from natural images.Therefore,it is not ideal to directly apply the object detection method for natural images to remote sensing imagery.Based on fully analyzing the properties and characteristics of optical remote sensing imagery,this paper proposes a novel method more suitable for remote sensing object detection to improve detection performance.The main research contents and innovations of this paper are as follows:(1)Aiming at the problem of high resolution and low detection speed of remote sensing imagery,this paper improves on the one-stage object detection method FCOS,which detects objects through an end-to-end network in a pixel-by-pixel way,and achieves both high detection accuracy and speed.This paper uses the simple and efficient Res Net-50 as the backbone to extract features.While improving the feature representation ability of the model,the detection speed of the network is accelerated.(2)Affected by the changing shooting environment and scenes,the background with complex distribution of remote sensing imagery is more difficult to distinguish from the objects,which causes great interference to object detection.Considering the problem of background with complex distribution and great interference to the objects,the structure tensor extractor is designed to efficiently extract the structural features and spatial details of objects in remote sensing imagery.At the same time,the structure tensor extractor achieves the purpose of screening out important object features in the background with complex distribution.Aiming at the problem that the small objects with dense arrangement in remote sensing imagery have few detailed features and are continuously lost in high-frequency downsampling operations of convolutional neural networks,a structure-guided feature transform module based on the attention mechanism is proposed.This module takes the structural features and spatial details obtained by the structure tensor extractor as learnable weights,and adaptively guides the structure and detail features to the end of the detection network to preserve and enhance the structural features of the object as much as possible,which effectively improves the detection performance of small objects with the dense arrangement.The overall accuracy is improved by 1.94%.(3)Affected by the shooting angle and height of the sensor,even similar objects in remote sensing imagery have various sizes and shapes,with great intra-class differences and a lack of significant feature information.Considering the problem of increasing the difficulty of detection due to the variety of object scales in remote sensing imagery,the multi-scale hybrid residual module constructs hierarchical residual connections with different convolution kernel sizes in the residual block to represent multi-scale information in a finer-grained manner and expand the network receptive field.The multi-scale feature expression ability of the network and the detection performance of multi-scale objects in remote sensing imagery are improved.The overall accuracy is improved by 4.12%.To evaluate the method in this paper,an ablation experiment is carried out to verify the effectiveness of each module,and then the method in this paper is compared with several prevalent and progressive object detection methods to verify the superiority of the method in this paper.Finally,the robustness of our method is verified by experiments on two datasets. |
PDF Full Text Request