Research On Remote Sensing Image Target Detection Based On Convolutional Neural Network And Attention Mechanism

Target detection is one of the basic tasks of computer vision to locate and classify the target of interest in the image.With the continuous maturity of remote sensing technology,although the number of remote sensing images has exploded,the utilization rate of remote sensing images is still very low.The utilization rate of aerospace remote sensing data is less than 5%,the utilization rate of aviation remote sensing data is less than 10%.How to get useful information from massive amounts of data?Therefore,it is urgent to develop intelligent image understanding and interpretation methods.Remote sensing images target detection technology is an important research direction for remote sensing image understanding and interpretation.With the development of computer technology and continuous improvement of deep learning theory,the performance of remote sensing images target detection technology has been greatly improved.However,because remote sensing images generally have uneven target distribution and many small targets,irregular target distribution is densely distributed,targets have different orientations and complex background environments,lead to target detection method based on the ordinary optical image in remote sensing image performance is poorer.Therefore,it has become a trend to study high-performance target detection models based on the specific characteristics of remote sensing images.Considering the above status of remote sensing images,this thesis focuses on solving the detection problems of small targets,dense targets and multi-directional targets,which based on characteristics of remote sensing images,convolutional neural network target detection theory and methods.Specific research is as follows:(1)Aiming at the problem that it is difficult for remote sensing images containing small targets with complex environmental background to carry out accurate target detection,a single-stage target detection model based on the fusion of attention and features is proposed.This model is an improvement of the one-stage target detection model SSD(Single Shot MultiBox Detector),which mainly consists of detection branch and attention branch.Firstly,the attention branch was added into the detection branch SSD,and the full convolutional network of the attention branch obtained the position characteristics of the target to be detected through per-pixel regression.Secondly,the feature fusion of the detection branch and the attention branch was carried out by the method of corresponding element addition to obtain the high-quality feature map with richer details and semantic information.Finally,Soft non-maximum suppression(Soft-NMS)was used for post-processing to further improve the accuracy of target detection.Experimental results show that the proposed model can significantly improve the detection accuracy and detection efficiency in the detection of small targets.(2)Aiming at the problem of the performance of existing target detection models in dense target detection,a remote sensing image target detection model based on attention and deformable convolution is proposed.Firstly,in the feature extraction stage,the feature pyramid module is designed,and the multi-layer features are fused by the way of subsampling and jumping connection,so that the feature map containing more information is obtained.Secondly,the attention mechanism is adopted to design attention module.The attention module is composed of multi-scale position attention and convolutional block attention module.By applying weights,they adaptively strengthen the location and channel features and reduce the interference of useless information.Finally,a deformable convolution module is used to adjust the receptive field adaptively by learning the offset of each pixel in the feature map,and effectively sample dense targets with different shapes and sizes,thereby improving the detection accuracy.Experiments show that,the proposed detection model not only detected dense targets well,but also have better performance than other comparison methods in the thesis.(3)Aiming at the problem of overlapping horizontal detection frames of multi-directional dense targets,this chapter proposes a remote sensing image target detection model based on directed bounding box prediction.The model first adds angle information to the RPN(Region Proposal Network),and realizes the detection of the oriented target by predicting the angle information.Secondly,reset the size of the anchor box to cover more remote sensing targets,and change the traditional intersection and union ratio algorithm for horizontal bounding boxes,and propose a new intersection and union ratio algorithm suitable for directed bounding boxes.Finally,the output image contains a directional bounding box,which contains less background information and fits the target under test well.Experimental results show that the proposed detection model can realize the detection of multi-directional dense targets,and also has higher detection accuracy.