Research On Small Target Detection Based On YOLOv5 Aerial Perspectiv

Unmanned Aerial Vehicle(UAV)flight operations have gradually become popular in recent years.However,due to the flight altitude of UAVs,the relevant development of UAVs combined with target detection algorithms has yet to be improved.This paper studies the small target detection algorithm and lightweight network based on the YOLOv5 algorithm for the characteristics of little feature information,large resolution of the live aerial view,limited memory,and computing power of UAV-embedded devices.The main research contents are as follows.(1)A target detection method based on feature fusion and contentaware feature reorganisation is proposed to fully use the limited feature information of targets to improve small target detection capability.Firstly,a K-Means-based optimised clustering module is utilised to reduce the randomness and human influence factors of the clustering algorithm;secondly,fusion operations are performed in the network using shallow small target features to achieve effective localisation of small targets;Finally,the content transformation and content-aware feature recombination module are used to transform the spatial information into channel information to realise downsampling operation and retain the complete feature information.Then the underlying content information is used to predict the adaptive and optimised recombination kernel at different locations.Then the feature recombination is carried out in the predefined nearby area to realise the upsampling operation.Experimental results show that the target detection method based on feature fusion and content-aware feature recombination can effectively improve the detection accuracy of small targets.(2)A sliding window cut-based aerial photograph small target detection method is proposed to solve the problem of poor aerial photograph detection.The training and testing datasets are pre-processed with sliding window cuts before training and prediction.Finally,all the sub-prediction results are stitched and drawn into the aerial image to improve the detection effect of the aerial image.The experimental results show that the sliding window cut-based aerial photograph small target detection method can significantly improve the detection accuracy of the network model for small targets.In particular,the detection effect for tiny targets is improved significantly.(3)To solve the problem of limited internal storage space and computational chips of UAV flight devices,two lightweight networks,Ghost Net and Shuffle Net,are used to compress the number of parameters of the aerial photography small target detection model with sliding window cut to obtain a lightweight detection network.The experimental results show that the improved lightweight network based on Shuffle Net can achieve target detection performance with fewer parameters,smaller models and higher accuracy.This paper investigates the small target detection algorithm and lightweight network based on the YOLOv5 algorithm.It uses the F1 and m AP values as evaluation indicators to verify the effectiveness of the improved method.After experimental verification and analysis,the small target detection method based on the YOLOv5 aerial photography perspective proposed in this paper can effectively improve the accuracy of the network model for small target detection and realise the lightweight design of the network.