摘要
针对织物疵点存在的种类多、密集度高、尺度小等检测难点,提出一种基于改进YOLOv5模型的织物疵点检测算法。首先,通过Kmeans++方法对所有真实框进行聚类,提高了模型训练时的收敛速度;其次,将Mish激活函数用于模型训练,提高了其非线性表达能力;再次,通过增加检测层提升了对多尺度目标的检测性能,并调整池化层位置提取多尺度的特征信息,提升了算法的鲁棒性及检测精度;最后,优化颈部网络结构,提升了算法的检测精度和速度。基于天池织物疵点数据集的实验结果表明:该算法的mAP达到了76.8%,相较基于原YOLOv5模型的织物疵点检测算法提升了7.7%,验证了该算法的有效性和鲁棒性。该算法在满足实时检测的要求下提高了检测精度,并优于其他主流目标检测算法,具有良好的应用前景。
In view of the difficulty in detecting fabric defects,i.e.,wide variety,high density and small scale,a fabric defect detection algorithm based on the improved YOLOv5 model was proposed.The algorithm firstly clustered all real frames with Kmeans++method and increased the convergence speed during model training.Secondly,the Mish activation function was used to train the model and enhanced its nonlinear expression ability.Thirdly,the detection performance of the model for multi-scale targets was improved by adding detection layers.The position of pooling layer was adjusted to extract multi-scale feature information and improve the robustness and detection accuracy of the algorithm.Finally,the neck network structure was optimized to improve the detection accuracy and speed of the algorithm.The experimental results based on the Tianchi fabric defect data set showed that the mAP of the proposed algorithm can reach 76.8%,which was 7.7%higher than that of the original YOLOv5model,verifying the effectiveness and robustness of the proposed algorithm.This algorithm improves detection accuracy while meeting the requirements of real-time detection and outperforms other mainstream target detection algorithms,so it has a good application prospect.
作者
郭波
吕文涛
余序宜
郭庆
陈亮亮
王成群
GUO Bo;LÜWentao;YU Xuyi;GUO Qing;CHEN Liangliang;WANG Chengqun(School of Information Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China;Zhejiang Technical Innovation Service Center,Hangzhou 310007,China;Department of Applied Engineering,Zhejiang Institute of Economics and Trade,Hangzhou 310018,China)
出处
《浙江理工大学学报（自然科学版）》
2022年第5期755-763,共9页
Journal of Zhejiang Sci-Tech University(Natural Sciences)
基金
国家自然科学基金项目(61601410)
浙江省科技厅重点研发计划项目(2021C01047,2022C01079)。
