Research On UAV Image Target Recognition Algorithm Based On Machine Learning

Nowadays,the market application prospect of unmanned aerial vehicle（UAV）is becoming more and more extensive.It is involved in most fields,such as monitoring,aerial photography,security,disaster relief and so on.With the increasing demand of UAV automation,object recognition technology based on computer vision has become one of the research hotspots.Due to aerial work,compared to the target recognition in other scenes,the environment of UAV vision is more complex,and the target feature information is less,with single features,small size,and other characteristics.And the shape of the non-rigid target changes when it moves.Improving the accuracy of target recognition is facing huge challenges in practical applications.To solve these problems,the major works are as follows:（1）There is relative motion between UAV and moving target during the flight,and the global coordinate system of the adjacent frame images often changes,which makes the background of the image of the UAV vision has rotation.At the same time,the non-rigid target also has the problem of deformation during moving.All these problems bring some difficulties to the UAV targets detection and recognition.Combined with the different advantages of visual salience algorithms in the time domain,frequency domain,and space domain,a visual salience image representation method based on the time,space and frequency domain is proposed in this paper,which can better highlight the shape and location information of potential targets.Otsu method segmentation algorithm is used to extract the candidate area of targets.Furthermore,HOG,PHOG and LBP operators are selected to extract the feature of potential target,and SVM classifier is used to recognize the final target.The experimental results on the Helicopter dataset show that the proposed method can effectively extract the candidate region of the target and detect the real target in the case of background rotation and target deformation caused by global motion,and the precision,recall and F₁ measure can reach more than 90%.（2）For the UAV images,the target size distribution is uneven,the number of small targets is large,and the targets are dense.All these lead to the problem of data loss when using Yolov3 network for target detection and recognition.This paper improves Yolov3network by changing the number of anchor points,multi-feature fusion,and introducing attention mechanism,and applies the improved Yolov3 network to UAV image target recognition.The experimental results on Visdrone2019 dataset show that the performance of proposed improved Yolo3 network for target detection and recognition is better than Yolov3,and it also has some advantages compared with other detection methods.（3）UAV is a mobile terminal device with limited storage space and computing capabilities.It requires high memory and large processing time if using a complex network structure for processing directly.Considering that lightweight convolutional neural networks,such as Mobilenetv3,Shufflenetv2 and Ghostnet,introduce effective convolution methods,this paper tries to use these networks as the backbone network for feature extraction,combining with the network Faster-rcnn and the network Yolov3-4 proposed in this paper for target recognition.and compares them with experiments.The experimental results on the Visdrone2019 dataset show that the network proposed in this paper is more suitable for target recognition in UAV vision than Faster-rcnn.At the same time,it is also verified that the feature extraction ability of the Mobilenetv3lightweight convolutional neural network is better than the other two lightweight convolutional neural networks.