Research On Image Detection And Recognition Algorithm Of Low-Altitude UAV Target Based On Deep Learning

With the development of automatic control technology,the more and more extensive application of civil UAV brings a lot of convenience to the society,but also brings serious threats to personal,social,military and other fields.Therefore,rapid and efficient detection of low-altitude UAV is the premise to deal with UAV threats.In this paper,the photoelectric detection means of infrared and visible light bands and deep learning technology are used to carry out the research on the image detection and recognition algorithm of low-altitude UAV.The main work contents are as follows:(1)Optimization of low-altitude UAV detection and recognition algorithm based on deep learning.First,the dual-band UAV data set was constructed by infrared and visible light acquisition equipment and type 3 civil UAV.Then,Faster R-CNN(Faster Region with Convolutional Neural Networks features),SSD(Single Shot Multi Box Detector)and YOLOv3(You Only Look Once version3)algorithms were compared and analyzed.Finally,a comparison experiment is carried out on the three algorithms based on the single band data set.The experimental results show that the performance of YOLOv3 algorithm is better than other algorithms.(2)The optimized YOLOv3 algorithm was proposed to improve the detection and recognition ability of low-altitude UAV targets.The algorithm optimizes and improves the residual network and prediction network for the characteristics of low altitude UAV.The experimental results showed that the m AP(mean Average Precision)value of optimized YOLOv3 algorithm was improved compared with that of YOLOv3 algorithm when the detection speed was reduced by 2 FPS(Frames Per Second).(3)In order to further improve the performance of the algorithm,three fusion detection algorithms were proposed by combining the fusion idea with detection and recognition based on dual-band UAV data.Specifically,the first algorithm is a first fusion and then detection algorithm based on deep learning.This algorithm first uses the image fusion algorithm to fuse dual-band image pairs,and then carries out model training and testing on the fused images.The results show that the algorithm is superior to the single-band detection and recognition algorithm in accuracy,but the detection speed is only 4 FPS.The second algorithm is a fusion and detection based on deep learning.The algorithm combines feature fusion with detection and recognition to construct a two-channel detection and recognition network.Compared with the first algorithm,the experimental results show that the m AP is 1% lower,but the detection speed reaches 27 FPS,with the best overall effect.The third algorithm is a first detection and then fusion algorithm based on deep learning.This algorithm performs decision-level fusion of the prediction results of a single band model,slightly improved its accuracy.By comprehensive comparison,the second algorithm works best.