基于正交优化策略的YOLO模型超参数优化方法  

Hyperparameter Optimization of YOLO Model Based on Orthogonal Optimization Strategy

作　　者：杨青华 杨观赐[1] 钟世昊 YANG Qing-hua;YANG Guan-ci;ZHONG Shi-hao(Key Laboratory of Advanced Manufacturing Technology of Ministry of Education,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学现代制造技术教育部重点实验室,贵阳550025

出　　处：《科学技术与工程》2025年第4期1573-1579,共7页Science Technology and Engineering

基　　金：国家自然科学基金(62163007,62373116);贵州省科技计划(黔科合平台人才[2020]6007-2,黔科合支撑[2021]一般439)。

摘　　要：为实现YOLO(you only look once)模型的超参数自动优化,提出基于正交优化策略的YOLO模型超参数优化方法(hyper-parameter optimization of YOLO model based on orthogonal optimization strategy,OOS)。首先基于统计学的正交试验原理,提出了种群的正交搜索方法与超参数贡献度分析策略,提高了算法的优化效率;然后,设计了均匀正交搜索策略和邻域正交搜索策略,以缓解YOLO模型陷入局部最优和早熟收敛问题。最后,在NWPU VHR-10和Pascal VOC两个目标检测数据集上,以YOLOv5、YOLOv5s-Transformer和YOLOv7为优化对象进行测试,测试结果表明,所提出的OOS超参数优化方法对于YOLO模型的识别精度均有所提升。在两个数据集上的平均识别精度mAP@0.5分别提升至93.94%、93.18%、93.45%以及85.81%、84.59%、90.62%;mAP@0.5-0.95提升至60.00%、60.08%、56.98%以及62.27%、58.89%、71.91%,可为目标检测模型的超参数智能优化提供一种新方法。In order to realize the automatic optimization of hyperparameters of YOLO model,the hyperparameter optimization of you only look once(YOLO)model based on orthogonal optimization strategy(OOS)was proposed.Firstly,based on the principle of statistical orthogonal test,the orthogonal search method of population and the hyperparameter contribution analysis strategy were proposed to improve the optimization efficiency of the algorithm.Then,the uniform orthogonal search strategy and the neighborhood orthogonal search strategy were designed to alleviate the problem of the YOLO model falling into the local optimum and premature convergence.Finally,YOLOv5,YOLOv5s-Transformer and YOLOv7 were used as optimization objects to test on two target detection datasets,NWPU VHR-10 and Pascal VOC.Test results show that the recognition accuracy of the YOLO model is improved by the OOS hyperparameter optimization method in all cases.The average recognition accuracy mAP@0.5 on two datasets is improved to 93.94%,93.18%,93.45%,and 85.81%,84.59%,89.96%.The mAP@0.5-0.95 is improved to 60.00%,60.08%,56.98%,and 62.27%,58.89%,70.77%.It can provide a new intelligent method for hyperparameter optimization of object detection model.

关 键 词：目标检测 超参数优化 则化策略 YOLO 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]