基于改进YOLOv5的视觉定位算法  

Visual positioning algorithm based on improved YOLOv5

作　　者：牛洪超 白松 胡晓兵[1,2] NIU Hong-chao;BAI Song;HU Xiao-bing(School of Mechanical Engineering,Sichuan University,Chengdu 610065,China;Yibin Institute of Industrial Technology,Sichuan University,Yibin 644000,China;Technology Management and Development Office,Luzhou Science and Technology Evaluation and Achievement Transformation Center,Luzhou 646000,China)

机构地区：[1]四川大学机械工程学院,四川成都610065 [2]四川大学宜宾产业技术研究院,四川宜宾644000 [3]泸州市科技评估和成果转化中心科技管理及开发室,四川泸州646000

出　　处：《计算机工程与设计》2022年第6期1620-1627,共8页Computer Engineering and Design

基　　金：中国制造2025四川省行动计划基金项目(2020QT-2020-jww-11-26);四川大学泸州市战略合作基金项目(2020CDLZ-1)。

摘　　要：针对传统视觉定位系统所存在的检测精度低等缺点,提出一种基于改进YOLOv5的自定义靶标视觉定位算法。为进一步增强网络的检测能力,分别融入具有多层感受野与细粒度的模块与改进的特征增强模块,利用Distance-IOU与Focal Loss改进损失函数;使用快速解码算法得到编码信息。实验结果表明,在自制数据集上,改进后的YOLOv5模型获得较好的平均准确精度得分与检测速度,满足实时性与准确性的需求,为视觉定位提供了一种的解决方案。Aiming at the disadvantages of traditional visual positioning systems such as low detection accuracy,a custom target visual positioning algorithm based on improved YOLOv5 was proposed.To further enhance the detection ability of the network,the multi-layer receptive field and fine-grained module and the improved feature enhancement module were integrated respectively for the network,and the Distance-IOU and Focal Loss were used to improve the loss function.The fast-decoding algorithm was used to get the coding information.Experimental results show that on the self-made data set,the improved YOLOv5 model achieves better average accuracy score and detection speed,meets the needs of real-time and accuracy,and provides a solution for visual positioning.

关 键 词：视觉定位 YOLOv5 感受野 细粒度 自定义靶标 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]