基于轻量化重构网络的表面缺陷视觉检测

基于深度学习的方法在某些工业产品的表面缺陷识别和分类方面表现出优异的性能, 然而大多数工业产品缺陷样本稀缺, 而且特征差异大, 导致这类需要大量缺陷样本训练的检测方法难以适用. 提出一种基于重构网络的无监督缺陷检测算法, 仅使用容易大量获得的无缺陷样本数据实现对异常缺陷的检测. 提出的算法包括两个阶段: 图像重构网络训练阶段和表面缺陷区域检测阶段. 训练阶段通过一种轻量化结构的全卷积自编码器设计重构网络, 仅使用少量正常样本进行训练, 使得重构网络能够生成无缺陷重构图像, 进一步提出一种结合结构性损失和L1损失的函数作为重构网络的损失函数, 解决自编码器检测算法对不规则纹理表面缺陷检测效果较差的问题; 缺陷检测阶段以重构图像与待测图像的残差作为缺陷的可能区域, 通过常规图像操作即可实现缺陷的定位. 对所提出的重构网络的无监督缺陷检测算法的网络结构、训练像素块大小、损失函数系数等影响因素进行了详细的实验分析, 并在多个缺陷图像样本集上与其他同类算法做了对比, 结果表明重构网络的无监督缺陷检测算法有较强的鲁棒性和准确性. 由于重构网络的无监督缺陷检测算法的轻量化结构, 检测1024 × 1024像素图像仅仅耗时2.82 ms, 适合工业在线检测.

Visual Inspection of Surface Defects Based on Lightweight Reconstruction Network

Deep learning-based methods show excellent performance in identifying and classifying surface defects of certain industrial products. However, most industrial product defect samples are scarce and feature differences are large, making it difficult to apply this type of detection method that requires a large number of defect samples. This paper proposes an image reconstruction-based unsupervised defect detection algorithm reconstruction network for defects detection, which uses only non-defective sample data that is easily available in large quantities to detect abnormal defects. The algorithm proposed in this paper includes two stages: image reconstruction network training stage and surface defect area detection stage. In the training phase, the reconstruction network is designed by a fully convolutional self-encoder with a lightweight structure, and only a small number of normal samples are used for training, so that the reconstruction network can generate defect-free reconstruction images, and a combination of structural loss and L1 is further proposed. The loss function is used as the loss function of the reconstructed network to solve the problem of poor detection of irregular texture surface defects by the self-encoder detection algorithm; the residual area of the reconstructed image and the image to be tested is used as a possible defect area in the defect detection stage. The final inspection result can be obtained through conventional image operations. In this paper, the network structure, training patch size, loss function coefficient and other influencing factors of the proposed reconstruction network for defects detection method are analyzed in detail, and compared with other similar algorithms on several defect image sample sets. The results show that reconstruction network for defects detection has strong robustness and accuracy. Due to the lightweight structure of reconstruction network for defects detection, it takes only 2.82 ms to detect 1024 × 1024 pixel images, which is suitable for industrial online detection.