A network lightweight algorithm based on improved Yolov4 was designed to address the issues of high computational complexity and low detection frame rate in object detection of surface vessels.Firstly,a DWG module is proposed,which is combined with Ghost convolution to form a new backbone for Yolov4,reducing the size of the network model.Secondly,SE attention is added in front of the neck network and the neck network is simplified to an FPN structure to improve the detection frame rate.Finally,Mish function is introduced to replace the activation function of the original network,and Focal Loss is used to optimize the loss function.The experimental results show that the improved algorithm reduces the number of parameters by 93.2%,computational complexity by 95.1%,and detection speed by 3.2 times compared to the original algorithm.In sum,the proposed algorithm can achieve real-time detection of surface vessels.
关键词
船只目标检测/轻量化网络/Yolov4算法/Ghost卷积/实时检测
Key words
vessel object detection/lightweight network/you only look once algorithm of version four/ghost convolution/real-time detection
维普
万方数据