Automatic Optimization of YOLOv3 Based on Particle Swarm Algorithm

YOLO (look once at a time) is a target detection method of deep learning, which has achieved great success in image classification and localization and has been widely used. YOLOv3 is the third version of YOLO. Compared with previous versions of software, YOLOv3 employs a better basic network (ResNet) and classifier in the network structure. It also uses multi-scale functions for object detection. For a neural network, the setting of hyperparameters is one of the important factors that determine the performance of the model. The hyperparameter of YOLOs is manually designed by experienced researchers. This process requires researchers to constantly turn parameters to enhance network performance and the experimental process is clueless. This chapter presents a method of particle swarm optimization combined with YOLOv3 for automatic parameter tuning. According to the optimization idea of particle swarm algorithm, the optimal hyperparameter is taken as the search target of particle swarm algorithm. Based on a series of comparative experiments, we can prove on the VOC dataset that the YOLOv3 network optimized by PSO achieves higher accuracy than the YOLOv3 network model before optimization.