In order to solve the slow response speed,high hardware resource occupation and other issues of automatic detection tasks under such complex conditions as stacked adhesions and debris interference,a part detection method based on lightweight YOLOv4 network is proposed.MobileNeXt is used to replace CSPDarkNet53 as backbone,and coordinated attention mechanism is added in each convolutional module to enhance the semantic expression ability of feature maps.A Fused-Sandglass module is proposed for plugging into shallow backbone,which can improve the inference speed of network.In the aspect of network training,the progressive training method and focal loss function are introduced to improve the training speed and effectively alleviate the imbalance between positive and negative samples.Experimental results show that the proposed method can maintain the accuracy similar to that of YOLOv4 network in 15 parts detection tasks,and the number of parameters is only 20％of it.The inference speed can reach 43.7 fps,which can meet the demand of actual production.
关键词
深度学习/协调注意力机制/零件检测/YOLOv4网络/MobileNeXt网络
Key words
deep learning/coordinating attention mechanism/part detection/YOLOv4 network/MobileNeXt network
维普
万方数据