Research On UAV Aerial Image Target Detection Based On Improved YOLOv5

UAV aerial monitoring has important application significance for urban planning management,traffic monitoring,environmental detection,agricultural production,search and rescue and other scenarios.With the development and progress of technology,deep learning algorithms have matured and have made great progress in UAV target detection.Aerial images have high resolution,large viewing angle,complex background,small and dense targets,and high requirements for target detection algorithms.General target detection algorithms based on deep learning are difficult to solve the above problems.Aiming at the shortcomings of the current algorithm in UAV aerial target detection,this paper improves the model feature extraction and fusion process on the basis of YOLOv5 s algorithm,and introduces parallel attention mechanism and efficient lightweight ideas.The main contents of this paper are as follows :1.Aiming at the problem that the UAV aerial image target is small,the feature is difficult to extract,and the target overlaps,a UAV aerial image target detection algorithm based on fine-grained features is designed.To retain and utilize more small target information in the process of model sampling,fine-grained feature extraction is carried out in the backbone network,and the path of small target detection head is added.The double downsampling is completed by interval sampling and convolution,and the improved module is connected by residual,and fused with the backbone network,and then the low-level network feature map is output.To retain effective information in the process of multi-scale feature transfer fusion,a multi-scale feature fusion method with adaptive feature enhancement is designed to make the feature map of the predicted small target have more effective semantic information.To improve the prior information obtained by the network and effectively use the information between adjacent prediction boxes,the K-DBSCAN clustering algorithm is designed to calculate the pre-selection box and the WSoft-NMS non-maximum suppression strategy to reduce the model miss rate.Through ablation experiments and comparative experiments,it is confirmed that the improved model is higher than the two-stage target detection algorithm and the anchor-free frame detection algorithm in recognition accuracy.2.Aiming at the problem of small target model with less feature information and too much calculation and parameter,a lightweight UAV aerial target detection algorithm based on attention mechanism is designed.Firstly,to make full use of the channel information and spatial information of small targets in the feature map,a small target attention module based on parallel mode is designed,so that the channel and spatial attention information in the feature map can not only use each other but also not directly affect each other,suppress useless information and highlight small target information.To reduce the amount of model parameters and calculation and maintain the feature extraction ability of the model,an efficient lightweight convolution module is designed to replace the original feature extraction module.The effectiveness of the proposed method is verified by ablation experiments and comparative experiments.It is also proved that increasing the input size appropriately can improve the target detection performance of UAV aerial images.