A Remote Sensing Object Detection Algorithm Based On Transformer And Relocation Non-maximum Suppression

As a hot point in the field of remote sensing,remote sensing images object detection plays an important role in military and aviation.With the development of remote sensing image acquisition technology,a large number of clear and highresolution remote sensing images are used for object detection.High-resolution remote sensing images can provide detailed information about the object and help to identify and locate the object better.However,the scale differences between objects in complex backgrounds make it difficult for various object detection algorithms to identify and locate remote sensing object accurately.Aiming at above problem,this paper investigates the feature extraction,feature fusion and non-maximum suppression algorithms.The main research contents of this paper are as follows:(1)Remote sensing object detection with complex backgrounds and multi-scale objects is still a challenge.Aiming at this challenge,we propose a Transformer-based remote sensing object detection algorithm(Trans Net).The backbone network of the algorithm is a Vision Transformer network termed Uniformer that takes advantage of its ability to model global information for the purpose of fully extracting object features under complex scenes.We further design an adaptive path aggregation network.In the designed network,CBAM(Convolutional Block Attention Module)is utilized to suppress background information in the fusion paths of different-level feature maps,and new paths are introduced to fuse same-scale feature maps for increasing the feature information of the feature maps.The designed network can provide more effective feature information and improve the feature representation capability.During model training,a migration learning approach is used to improve the convergence speed of the model.The experiments on Trans Net are conducted on RSOD,NWPU VHR-10 and DIOR datasets.Compared with the current popular object detection algorithms,Trans Net has better detection performance.(2)To address the problem that the localization accuracy of the optimal bounding box selected by the traditional NMS algorithm is low.We propose a novel Relocation non-maximum suppression(R-NMS)algorithm for remote sensing object detection and design a new box distance measurement method to replace Intersection over Union to measure the distance between two bounding boxes.First,the bounding box with the highest score of classification confidence in the bounding boxes set is selected as the optimal bounding box.Then,the location information of the bounding boxes around the optimal bounding box is obtained by the new box distance measurement method.Finally,the location information is used to relocate the optimal bounding box to obtain the new optimal bounding box.To prove the effectiveness of R-NMS algorithm,this paper uses R-NMS algorithm as the post-processing of Trans Net and conducts experiments to compare with the traditional NMS algorithm and Soft-NMS algorithm.The experimental results show that Trans Net combined with R-NMS algorithm has good detection performance.Figure [20] Table [17] Reference [61]...