Detection Of Small Objects In Remote Sensing Images Based On Cascade Attention And Distance Constrained Regression

Remote sensing image target detection is one of the hot topics in the field of computer vision and remote sensing.Its research is of great significance for both military and civilian use.At present,when performing tasks related to remote sensing image target detection,the mainstream methods have the following problems : small targets in remote sensing images are not obvious enough,resulting in easy loss of detected target features;the small target feature expression is not significant,and the effective feature extraction is difficult;it is difficult to detect the prediction box regression of remote sensing targets.In view of the above problems,this dissertation proposes a series of corresponding solutions.The specific work is divided into three parts:(1)Aiming at the problem of feature loss when the detection method extracts the features of small targets in remote sensing images,this dissertation re-evaluates the processing methods of different scale feature maps in the feature fusion stage,and designs a remote sensing image detection method based on scale hierarchical feature pyramid.The method includes a high-level semantic information activation module and a low-level effective feature perception module.According to the different scales of the image in the feature extraction structure,it is processed separately to learn effective small target features.The m AP of the detection method in the NWPU VHR-10 dataset is 3.2 % higher than that of the original method without adding the scale-based hierarchical feature pyramid.(2)In order to solve the problem of difficult effective feature extraction of remote sensing images,a remote sensing image target detection method based on cascade attention mechanism is proposed.This method introduces Non-Local Attention(NLA),and forms a cascade structure with the low-level effective feature perception module to effectively extract the global features that are difficult to learn by conventional methods.In this structure,a multi-scale feature fusion operation is designed to effectively integrate the location and semantic information contained in different scale feature maps.Compared with the original method,the m AP of the detection method in the NWPU VHR-10 dataset is increased by 4.6 %,and the m AP in the RSOD dataset is increased by 2.7 %.(3)Aiming at the difficulty of detection frame regression caused by the complex background of small targets in remote sensing images,a remote sensing image target detection method based on distance constrained regression(DCCN)is proposed based on the anchor-free target detection method FCOS.DCCN redesigns the regression evaluation conditions according to the distance between the predicted sample box and the real sample box,and forms a new regression method,which optimizes the regression method of the network and avoids introducing additional parameters.The m AP of this detection method in NWPU VHR-10 dataset is 5.4 % higher than that of the original method.In addition,the m AP of the remote sensing small target image detection method based on the cascade attention and distance constraint regression formed by all the structures proposed in this dissertation reaches 92.3 % on the NWPU VHR-10 dataset,which is increased by 6.3 %,and the m AP in the RSOD dataset reaches 95.9 %,and the detection accuracy of the small target class is higher.Figure [19] Table [15] Reference [72]...