Research On Improved Detection Network Of Aerial Target Based On Deep Learning

There are relatively few studies on air target detection,and compared with ground targets,the situation when detecting air targets is more complicated.This paper focuses on the task of aerial target detection,deeply studies the existing target detection technology,analyzes the characteristics,advantages and disadvantages of various detection network optimization techniques,and proposes an improved structure based on the YOLOv4 network,focusing on the task of aerial target detection.The network improves the detection effect of aerial targets without increasing the computing time.In order to solve the problem of too few available samples,this paper uses the data augmentation method to generate a set of selfmade aerial target datasets with diverse target poses and changing environments,and train the network on the expanded datasets.The results of the test are:Compared with the original YOLOv4 network and other common target detection models,the optimized aerial target detection network in this paper has slightly improved the detection accuracy while basically maintaining the operation speed.It is relatively most obvious and has certain practical application value.The main contents of this article are as follows:The first part introduces the existing research results of object detection,including non-deep learning methods in the early years and deep learning methods in the later period.The second part introduces the principle and structure of mainstream target detection methods.Through a set of comparative experiments,it is confirmed that the YOLOv4 network has a good detection effect in the field of aerial target detection.Next,through in-depth analysis of the structure of the YOLO series network and commonly used optimization network methods,a plan for improving the network is drawn up from the following five aspects: modifying the CSP structure and head structure in the network,adding attention mechanisms,adding regularization processing items,Network structured pruning and changing loss function.The third part evaluates the improved plans by designing multiple sets of comparative experiments,abandons the invalid plans,and obtains the finalized improved YOLO network.First,modify the CSP structure of the backbone network,add a halved residual module to increase the network depth,and modify the spatial pyramid structure to make the network more suitable for the task of aerial target detection;second,add an attention mechanism to the network,Increase the overall efficiency of recognition;finally,the CIo U loss with the best comprehensive adaptability is selected as the loss function of the network.In addition,in order to facilitate experiments and tests,this paper also selfmade a small-sample aerial target data set for aerial target detection tasks,and carried out secondary development based on the open source image annotation software Labelimg,and obtained a multi-functional target detection experimental software platform.The self-made data set in this paper contains more than ten aerial target categories,and the total number of images exceeds 100,000.Various data augmentation methods are used to imitate the actual aerial target detection scenario;the multi-functional software platform integrates image annotation,data augmentation and network training Integrated with the test function,it can improve the efficiency of related work.In the experimental part,this paper chooses to test on the self-made small sample aerial target dataset.In the comprehensive performance test,the finalized improved YOLO network is compared with the original YOLOv4 network.The test results show that the accuracy of the model is slightly improved,and the inference speed is basically the same.