Optimization Algorithm for Real-Time Detection of Small Targets Based on YOLOv4-Tiny Structure
Targeting the practical application requirements of real-time detection of small targets,the Yolov4-Tiny net-work is taken as the basic framework,the ECANet is used to redesign the Bneck structure of MobileNetV3 and the main feature extraction network CSPDarkNet53-Tiny is replaced,to improve the depth and detection speed of the model;the depth and detection speed of the model is enhanced by adding the SPPCSPC module to the output interface of its back-bone network after the SPPCSPC module and replacing the feature pyramid(FPN)with a path aggregation network(PAN)to enhance the sensory field of the model and aggregate multi-region contextual information to get more adequate semantic and location information for each feature layer;and the CBAM attention mechanism is incorporated after the Head to enhance the useful information and inhibit the useless information to improve the detection accuracy of the model.The proposed algorithm is verified by real-time monitoring of mask wearing state,and the experimental results show that compared with YOLOv4-Tiny algorithm,the average accuracy of the proposed algorithm improves by 4.13%to 91.84%,and the FPS improves by 4.4 frame/s to 89.5 frame/s,which meets the real-time requirements of mask wearing state detection.
万方数据
维普
