Multi-Object Recognition Based on Improved YOLOv4
Xiaomin Huang, Shaolin Hu, Qiliang Guo · 2021 CAA Symposium on Fault Detection, Supervision, and Safety for Technical Processes (SAFEPROCESS) · 2021
When traditional deep learning model is used for multiple object recognition on the road, it has some problems such as large amount of calculation and poor timeliness. In order to solve these problems, this paper proposes a lightweight improvement strategy based on the YOLOv4 algorithm. First, in basic network architecture, the lightweight network MobileNetv2 is used to replace CSPDarknet53, which effectively reduces the amount of parameter extraction; then, in the deep network structure, the ordinary convolution layer is replaced with the deepthwise separable convolution, which further reduces the network parameters. Finally, we realized the lightweight YOLOv4 network. The multi-object recognition data set is used to train and test YOLOv4 and lightweight YOLOv4 respectively. The experimental results show that the mAP of the improved YOLOv4 network is 85.12%, which can effectively achieve the multi-object recognition task on the road. At the same time, compared with the YOLOv4, the detection speed of the improved network is increased by 75%, and the parameters are significantly reduced, thus saving about 83% of the storage space. The improved YOLOv4 has better application prospects in multi-object recognition task on the road.