Research On Target Detection Algorithm Of UAV Images Based On Deep Learnin

With the continuous development of artificial intelligence technology,smart devices combined with deep learning technology are widely used in many fields.Unmanned Aerial Vehicle(UAV),as a popular smart device,combined with object detection technology has important application value in fields such as traffic monitoring,forest security,search and rescue.The UAV aerial images are characterized by large differences in object size and a large number of small objects,and the object detection models in general scenes are unable to perform the task of object detection in UAV scenes better.Due to the problem of large differences in object size and the number of small objects in UAV aerial images,object detection models for generic scenes cannot perform the task of object detection in UAV scenes.Therefore,it is important to research the object detection algorithms for visible and infrared images in UAV aerial images.A lightweight feature multiplexing module and a multi-scale feature fusion network based on feature enhancement are proposed to solve the problem of high number of small targets and obscure features in visible images.Firstly,a residual structure-based feature reuse module is designed to improve the efficiency of the object features in the backbone network,while Atrous Spatial Pyramid Pooling is combined with an attention mechanism to enhance the feature representation in the multi-scale feature fusion network.To address the problems of large number of parameters and unbalanced distribution of positive and negative samples of the existing algorithm,the proposed feature reuse module and multi-scale feature fusion network are integrated into the YOLOX algorithm,and the YOLOX algorithm is improved using Ghost module,CDIo U Loss,Focal Loss and other methods.The experimental results show that the improved YOLOX algorithm improves the detection accuracy by 5.08% and reduces the inference time by 2.21 ms on the Vis Drone2019 dataset.A lightweight UAV infrared image target detection algorithm based on multi-feature fusion is proposed for problems such as sparse infrared image texture features and slow algorithm inference.Lightweight feature extraction network with multi-feature fusion network with adequate feature extraction and classification and regression of features using anchor free frame methods.A comparison of the proposed algorithm with current mainstream algorithms on the HIT-UAV dataset is carried out to verify the advancement of the algorithm.