Research On Multi-Angle And Small Object Detection Algorithm In High-Resolution Remote Sensing Image

Object detection task which is one of the core issues in the field of computer vision consists of two main tasks: predicting the position and identifying the category of the object.Because the visual elements such as the appearance and shape of various kinds of objects are different,and there are objective factors interfering with the imaging process,object detection has always been a challenging task.With the rapid development of deep learning technology and the introduction of various classical neural networks,the performance of object detector based on deep learning has become faster and more accurate.With the development of remote sensing technology,it is possible to obtain ultra-high resolution remote sensing images,which contain more visual details.However,the current high-performance detection algorithms in natural perspective images cannot achieve ideal performance in object detection tasks in remote sensing images.It is mainly caused by some characteristics of remote sen sing images,such as large image size,complex background information,multi-angle distribution of targets,small target size,etc.These characteristics bring great diff iculties and challenges to the remote sensing image target detection task.This paper further analyzes these characteristics and summarizes them into two difficult problems of target detection in remote sensing images: accurate angle fitting of multi-angle targets and high-precision detection of small targets.Therefore,this paper focuses on the above issues,and the specific work is as follows:(1)Aiming at the problem of accurate angle fitting of multi-angle targets in remote sensing image object detection,a multi-angle object detection algorithm based on bidirectional attenuation Io U(Intersection-over-Union)loss is designed.Specifically,this function simulates Skew-Io U by Gaussian product,and attenuates the product from two directions according to the deviation of the predicted position.The bidirectional attenuation Io U loss has stronger trend-level alignment with Skew-Io U,thanks to its ability to reflect the Skew-Io U change caused by position deviation.The experiment on DOTAv1.0 dataset shows that this method achieves higher detection accuracy than mainstream methods under various loss forms and different accuracy conditions.(2)Aiming at the problem of high-precision detection of small objects in remote sensing image object detection,this paper designs a new detector for small objects based on YOLOX target detector.This detector realizes high-precision detection of small objects in remote sensing image through the unique multi-feature auxiliary structure.Specifically,a multi-feature auxiliary structure is designed,which obtains the auxiliary features of the object through a key point subnet to enhance features of the major network,and then uses enhanced features to improve the classification and regression ability of the major network;And then,the bidirectional attenuation Io U loss is used as the regression loss function,and the size of the feature map sent into the head is modified according to the size characteristics of small objects.Experiments on DOTAv1.0 dataset show that each improvement can effectively improve the detection accuracy of multi-angle small objects in remote sensing images.And the new multi-angle small object detector achieves the highest mean average precision in comparison with many mainstream detectors.