Research On Multi-angle Target And Fine Recognition Method Of High-resolution Remote Sensing Image Based On Transformer

Research into remote sensing images has gradually accelerated with the launch of artificial satellites into space and the development of artificial intelligence in science and technology.More people were utilizing remote sensing image processing in all aspects of their lives and work as a result of its development.The domain of remote sensing image technology has brought fine-definition and massive image data to the field of remote sensing image processing.Better data sets and advanced basic theories have created more possibilities for the field of remote sensing image processing.By the location and classification of objects to achieve aim,remote sensing image target detection and recognition technology is more commonly employed in agriculture and forestry,meteorology,topographic survey,military survey,etc.The target’s characteristics are more obvious from the top-down perspective used to capture remote sensing images,and the primary qualities are: a large-scale difference,a big aspect ratio,more obvious detail features,a dense arrangement,and variable orientation.Due to the significant difference in target scale,the remote sensing image target detection and recognition task cannot be directly achieved by utilizing the method in the natural scene.But,the detailed features are more obvious,offering an appropriate dataset for the fine recognition work.Thus,it is essential to provide an efficient solution to the problem of fine target recognition in remote sensing images.This paper proposes a high-resolution remote sensing image multi-angle target fine detection method based on Transformer to solve the mentioned problems.The problem of roteted detection and fine recognition in remote sensing images will be resolved in this paper from the perspectives of rotation detection and fine recognition,including:(1)For remote sensing image targets are arranged dense and changeable,which leads to difficulty in testing,designing the TRANSFORMER rotating target detection method based on angle classification: First,a remote sensing image rotating target detection detection method based on Transformer and angle classification is proposed,which increases the accuracy of remote sensing image rotation object detection task.This method addresses the problem of difficult target detection with variable direction.At the same time,the boundary discontinuity problem that occurred when trying to finish this task in the past using the angle regression method is solved.Second,to strengthen the extraction of local information from the model and improve detection and recognition performance on densely arranged objects,the extrusion and excitation module(SE block)is added to the encoder of Transformer.It is completed to solve the problems of false detection caused by dense target arrangement and overlapping of the generated detection frame.This method achieves an mAP value of81.33% on the DOTA dataset.(2)For remote sensing image target scale differences and large vertical and horizontal ratio,it has led to the problem of low accuracy of fine identification tasks:First,a single-level feature pyramid structure is created to fuse features of different scales and improve the detection performance performance of the model on targets with large scale differences and large aspect ratios.It addresses the problem of poor detection performance on targets with large scale differences and large aspect ratios.Second,deep convolution is added to the fully connected layer in the transformer decoder so that the model pays more attention to the information on local features,enhancing the accuracy of the remote sensing image fine recognition task.Third,deep convolution is added to the fully connected layer in the transformer decoder to pay more attention to the detailed features of the target.Last but not least,the weight constraint module is made to restrict the attention calculation from the global to the local scope,accelerate the model’s convergence and reducing its computational cost.This method achieves an an mAP value of 21.29% on the FAIR1M dataset.