Research On Small Object Detection Algorithm In Aerial Images Based On Improved YOLOX

In recent years,UAV technology has developed rapidly,and the images obtained by computer vision technology acting on UAV aerial photography can meet a variety of demand scenarios,such as emergency disaster relief and power inspection.However,the detection accuracy of the general deep learning object detection algorithm in aerial images is not satisfactory due to the problems of small objects of aerial images themselves such as few pixels and complex background.Therefore,this thesis begins to study and improve the general object detection algorithm from the aspects of feature extraction network and multi-scale feature fusion,so as to improve the detection accuracy of small objects in aerial images.The main work of this thesis includes:(1)Collecting current aerial image datasets DOTA and NWPU VHR-10 datasets,screening out six types of objects from them.Experimentally,it is found that the resolution span of images in the DOTA dataset is large and cannot be directly used for subsequent training,so it is processed,and finally the two datasets are merged to form the small object detection dataset of aerial images in this thesis.(2)By comparing the accuracy and efficiency of common object detection algorithms,the YOLOX algorithm is selected as the main architecture of the network for this experiment.For the problem of complex background in aerial images,this thesis proposes to embed a double ECA(Efficient Channel Attention Module)attention mechanism in the feature extraction network,which can weaken the weight of irrelevant background features and enhance the network’s attention to small object feature information,thus improving the accuracy of small object detection in aerial images.(3)In order to solve the problem of small object pixel in aerial images,this thesis analyzes the algorithm and concludes that although shallow features in the Neck structure contain more location information of small objects,but less semantic information,deep features have more semantic information.This thesis proposes to incorporate an improved multi-scale feature fusion module MRAE(Multi Resolution Attention Extractor)into the Neck structure of YOLOX to effectively fuse feature maps at different levels by assigning appropriate weight values according to the requirements of small object detection.It has been verified that the proposed improvement improves the detection accuracy of small objects,and the precision of the improved algorithm is 86.68%,which is 2.52% higher than that of the original YOLOX.In summary,this thesis takes the detection of small object in aerial images as a research task,and designs the small object detection algorithm of aerial images based on improved YOLOX,which effectively improves the precision of small object detection and provides a scheme for the progress in the field of small object detection.