Abstract: To detect the surface defects of product in the production line,a detection method based on Faster-Regions with Convolutional Neural Network (Faster R-CNN) was proposed,which was used to distinguish defects and mark its location.At preprocessing stage,a regional planning means was put forward to roughly cut out principal part of the defect for avoiding the generation of mass redundant windows and improving detection's speed and accuracy.Data expansion and K-fold cross validation were also used to increase the number of images and improve the robustness of the algorithm.To promote accuracy of position detection and identification,the sparse filtering was blended in convolution neural networks to extract deep characteristics as inputs for Faster R-CNN.Typical defects of the container mouth in filling production line were used as experimental object to verify the feasibility of the proposed algorithm.

Key words: surface defects detection, faster regions with convolutional neural network, location detection, sparse filtering, production process monitoring

摘要： 为检测生产线中产品的表面缺陷,提出一种基于卷积神经网络快速区域标定(Faster R-CNN)的缺陷检测方法,用于识别缺陷类型并标记出缺陷位置。预处理阶段提出区域规划方法粗略裁剪出缺陷主体,以避免产生大量冗余窗口,从而提升检测速度和精度。所提算法结合数据扩充方法增加了图像数量,通过划分K折交叉验证数据集改善了算法的鲁棒性;同时,将稀疏滤波思想融入卷积神经网络,提取双重深度特征作为Faster R-CNN的输入,提升了Faster R-CNN位置检测和识别的精度。通过油辣椒灌装生产线的封盖面典型缺陷检测验证了所提方法的可行性。

关键词: 表面缺陷检测, 卷积神经网络快速区域标定, 位置检测, 稀疏滤波, 生产过程监控