Research On Object Detection Methods For Optical Remote Sensing Images

At present,optical remote sensing image target detection technology has a wide range of applications in military,environment,urban planning,agriculture,forestry and other fields.Compared with natural images,remote sensing images have the characteristics of large and dense distribution of small objects,diverse directions,complex backgrounds and wide range of object scales,which bring certain challenges to the application of general object detection algorithms on remote sensing images.Many targeted improved algorithms have been proposed for this problem,but most of them only start from one or two difficult problems in remote sensing images,and lack a certain degree of comprehension,and their detection performance is difficult to reach a higher level than that of ordinary images.In order to detect objects in optical remote sensing images more quickly and effectively,this thesis designs and improves an algorithm with more comprehensive detection performance according to the characteristics of remote sensing images,such as many small targets,diverse directions,complex backgrounds and large scale differences.The main work is as follows:(1)Considering the characteristics of many small targets and dense distribution in remote sensing images,several mainstream target detection algorithms were compared on the same dataset(DOIR)for detection accuracy.According to the experimental results,YOLOv5 algorithm with better comprehensive detection performance,especially the best detection performance for small objects,was used as the benchmark algorithm.(2)Aiming at the problem of arbitrary target direction in remote sensing images,this thesis improves a object detection method of YOLOv5 rotation box.The two-dimensional Gaussian distribution is used to represent the five-parameter rotation box,and the metric between the corresponding two Gaussian distributions(KLD)is used to approximate the Io U of the two rotation boxes in the original image.Moreover,the label assignment strategy and loss function are redesigned.This method can better measure the gap between different prediction boxes and the same real box,and only adds a small amount of additional parameters.(3)Aiming at the complex background and large scale difference of optical remote sensing images,this thesis designs a pixel-based multi-dimensional attention mechanism.The pixel attention module is used as the main part,and the pixel-level feature learning is carried out in three different sizes of local scopes according to different sizes of objects with a single pixel as the center.Each local scope is scanned in four directions to convey local regional information,avoiding the interference of global context information and paying more attention to relevant regions.The existing channel attention module is then applied to determine the importance between different channels.The combination of the two attention modules can better highlight the foreground and weaken the background,and avoid the blurring of object boundaries.In addition,adding the multi-dimensional attention module to the seven positions of the enhanced feature extraction network of YOLOv5 can further improve the feature representation ability of objects of different sizes.Finally,this thesis design the experiments needed for the improved algorithm,and conduct a series of comparison and ablation experiments on the algorithm.The experimental results show that the m AP of the improved rotating YOLOv5 algorithm in this thesis can reach 75.11 on the DOTA-v1.0 test dataset.After adding the pixel-based multi-dimensional attention mechanism,the m AP is increased by 2.2 percentage points to 77.31,which has a great accuracy advantage compared with other remote sensing image object detection algorithms at the same time.