基于多尺度融合的一次性竹筷表面缺陷检测方法  

Surface defect detection method of disposable bamboo chopsticksbased on multi⁃scale fusion

作　　者：周宇博 沈岳[1] 匡迎春[1] Zhou Yubo;Shen Yue;Kuang Yingchun(College of Information and Intelligence,Hunan Agricultural University,Changsha 410128,China)

机构地区：[1]湖南农业大学信息与智能科学技术学院,长沙410128

出　　处：《现代计算机》2024年第5期61-66,共6页Modern Computer

摘　　要：表面缺陷检测识别是产品质量控制过程中的一项重要任务,针对目前一次性竹筷生产检测智能化程度低、误检率高等问题,提出了一种识别缺陷类别和定位缺陷的跨尺度加权特征融合网络。由于缺陷边缘不明显、尺寸小、背景纹理干扰等问题,提出改进的Retinex图像增强方法。然而数据生成过程昂贵且带有缺陷样本的数据很少出现,使用大量正常一次性竹筷样本进行特征提取,提高表面缺陷识别任务的泛化能力,并将逆残差架构与坐标注意力机制(CA)相结合,以增强多尺寸缺陷检测的鲁棒性。实验结果表明,所提出的方法可有效提升一次性竹筷检测识别的性能,检测准确率可达92.6%,满足实际生产中的需求。Surface defect detection and identification is an important task in the product quality control process.In view of the current problems of low intelligence and high false detection rate in disposable bamboo chopsticks production inspection,a crossscale weighted feature for identifying defect categories and locating defects is proposed.Converged Network.Due to problems such as unclear defect edges,small size,and background texture interference,an improved Retinex image enhancement method is proposed.However,the data generation process is expensive and data with defective samples rarely appears.A large number of normal disposable bamboo chopsticks samples are used for feature extraction to improve the generalization ability of surface defect recognition tasks,and the inverse residual architecture is combined with the Coordinate Attention mechanism(CA)are combined to enhance the robustness of multi-size defect detection.Experimental results show that the proposed method can effectively improve the performance of disposable bamboo chopsticks detection and recognition,and the detection accuracy can reach 92.6%,meeting the needs of actual production.

关 键 词：缺陷检测 机器视觉 注意力机制 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TS972.23[自动化与计算机技术—计算机科学与技术]