基于改进的YOLOv5色纱颜色识别方法研究  

Research of color yarn color recognition method based on improved YOLOv5

作　　者：陆鸿路 陈玉洁[1] 许高平 LU Hongu;CHEN Yujie;XU Gaoping(Donghua University,Shanghai,201620,China)

机构地区：[1]东华大学,上海201620

出　　处：《棉纺织技术》2024年第7期47-53,共7页Cotton Textile Technology

基　　金：国家重点研发计划项目(2022YFB4700603)。

摘　　要：针对色纱目标小、相近颜色区别不明显导致难以实时精准识别色纱颜色的问题,提出一种基于改进的YOLOv5模型色纱颜色高效识别方法。以YOLOv5网络为基础,在Backbone结构中引入倒残差结构,结合C3模块深度可分离卷积提升网络计算速度;利用加权特征融合机制,在保留丰富的位置和颜色信息基础上提升模型训练效率,同时引入注意力机制以提高网络的收敛性和鲁棒性。试验表明:优化后YOLOv5模型的平均精度均值达到了99.5%,与原始YOLOv5模型相比,颜色识别速度提高11.5%,模型参数量减少35%,迭代次数减少30%,网络模型收敛性更好。优化后YOLOv5模型可满足色织生产中色纱颜色快速准确识别的要求。Aiming at the problem that it was difficult to accurately identify the color of color yarn in real-time due to the small target of the color yarn and the inconspicuous difference between similar colors,an efficient color identification method of color yarn based on improved YOLOv5 model was proposed.The model was based on the YOLOv5 network,and the inverted residual structure was introduced into the Backbone structure,combined with the depth separable convolution of the C3 module to improve the computational speed of the network.The weighted feature fusion mechanism was used to improve the training efficiency of the model based on retaining the rich positional and color information,and the attention mechanism was also introduced to improve the convergence and robustness of the network.The experiments showed that the mAP value of the improved YOLOv5 model was reached 99.5%,the color identification speed was improved by 11.5% compared with the original YOLOv5 model,the amount of model parameters was reduced by 35%,the number of iterations was reduced by 30%,and the network model had better convergence.The improved YOLOv5 model could meet the requirements of fast and accurate color identification of color yarns in color weaving production.

关 键 词：色纱颜色识别 深度学习 YOLOv5 倒残差结构 注意力机制 加权特征融合 

分 类 号：TS101.8[轻工技术与工程—纺织工程]