Research On Aircraft Detection Method Based On Remote Sensing Images

As aircraft play a vital role in the civil and military fields,with the rapid development of high-resolution satellite remote sensing technology,more and more attention has been paid to aircraft detection in remote sensing images.With the development of image processing and deep learning,the effectiveness of aircraft detection has been effectively improved in recent years.However,as the remote sensing image covers a wide area,the background is complex and the targets are numerous,which affects the extraction of aircraft feature information,some obvious or structurally similar objects in the image are easy to be misidentified as aircraft.On the other hand,the Smooth L1 loss function is generally used to calculate the boundary box loss in the detection task,while the evaluation standard of the detection box is IoU.The correlation between the two is weak,which makes the generated detection box and the target have a large deviation.Moreover,in the remote sensing image,the aircraft is dense,and the traditional NMS algorithm is prone to cause the missing detection.To solve the above problems,this paper proposes to use deconvolution and positional attention module to extract the structural features of aircraft to reduce error detection.Taking GIoU as the evaluation standard of detection box,a loss function of boundary box based on GIoU is proposed to unify the two,so as to get the best detection box.At the same time,a GIoU-soft NMS algorithm(GIoU-soft NMS)is proposed to reduce the missed detection of clustering targets.The main research contents are as follows:Firstly,the deconvolution module is proposed to improve the generation of feature pyramid,aiming at the problem that the prominent object in the image is easy to be misidentified as an aircraft.As the remote sensing image is captured from top to bottom,the structure information of the aircraft is obvious.In order to extract the structural features of aircraft,the deconvolution module is used to increase its ability to extract the structural information of aircraft,so as to improve the expression ability of aircraft features and reduce error detection.Secondly,in view of the similar structure of error detection,this paper use position attention mechanism to capture spatial dependencies between any two locations of the feature map.It can integrate similar features from the global vision adaptively,highlight the aircraft feature and suppress the background information,enhanced feature representation ability,so as to distinguish between similar to aircraft structure of target in complex background.Thirdly,to solve the problem of weak correlation between Smooth L1 loss function and IoU as evaluation criteria for detection box,GIoU was used as evaluation criteria for detection box,and a boundary box regression loss function based on GIoU was proposed,which achieved the consistency between evaluation criteria and loss function and made the generated detection box closer to the target.Finally,given that when detecting the clustered aircraft the traditional NMS algorithm is easy to miss the target,a soft decision NMS algorithm based on GIoU is proposed to solve this problem.It avoids the situation that the detection box is deleted directly in the case of high overlap rate,which leads to the omission of detection.Combined with the above methods,this paper proposes an aircraft detection model based on structural features and GIoU,which not only enhances the ability to extract aircraft structural features,but also improves the ability to generate detection boxes.The above method is verified on two data sets of DOTA and DIOR,and the results show that the model can effectively reduce the error detection in aircraft detection,optimize the generation of detection box,reduce the missed detection,and improve the aircraft detection accuracy.