文献类型： 会议论文

第一作者： Bo Zhang

作者： Bo Zhang ¹ ; Wenbai Chen ¹ ; Xiaohao Wang ¹ ; Chunjiang Zhao ² ;

作者机构： 1.School of Automation, Beijing Information Science and Technology University

2.Beijing Academy of Agriculture and Forestry Sciences

关键词： Object detection;Lightweight network;3C industry

会议名称： CAAI International Conference on Artificial Intelligence

主办单位：

页码： 330-342

摘要： The object detection algorithm for smartphone assembly parts in the 3C (Computer, Communication, Consumer Electronics) scene occupies many system computing resources, and the flexible targets and small-scale heterogeneous components in the background lead to problems such as low detection accuracy. Based on the YOLOv5n network, we propose a lightweight and high-precision network-YOLOv5n-GTA. First, we replace the ordinary convolution module with the Ghost convolution module in the backbone network and neck, thereby significantly reducing the parameter scale in the network. Second, we add a Transformer module at the end of the backbone network to enhance the feature expression ability of the network and further improve the modeling and representation ability of the backbone network. Then, We use the meta-ACON (Activate or not) activation function to dynamically learn the activation function's linearity and solve the neuron necrosis problem by controlling the nonlinearity of each layer of the network. Experimental results show that our method outperforms other excellent detection algorithms in inference speed and model size evaluation indicators. We compress model parameters to 63% of YOLOv5n and model size to 75% of YOLOv5n at a speed of 24.5 ms per image on GPU.

分类号： tp18-53