鲁棒主成分分析的铝箔表面缺陷检测方法

为了准确检测铝箔表面的穿孔、污点、亮斑和刮痕等各种缺陷，提出了一种基于低秩稀疏分解的铝箔图像表面缺陷检测方法。铝箔材料生产过程中表面出现缺陷的概率较小，同时一幅铝箔图像中缺陷占整幅图像的比例较小，即铝箔图像背景之间是线性相关的，可近似视为处于同一低秩子空间中，同时图像表面缺陷是近似稀疏的。采用RPCA(Robust Principal Component Analysis)算法对铝箔图像序列组成的观测数据矩阵进行低秩稀疏分解，得到低秩的背景图像和稀疏的缺陷图像。分别对单幅铝箔图像以及由多幅铝箔图像组成的图像序列进行低秩稀疏分解实验，在铝箔图像表面缺陷检测应用中验证所提方法的有效性。实验结果表明，提出方法检测到的缺陷清晰、完整，处理一幅大小为880×540的铝箔图像平均耗时不超过0.7秒，能够实现铝箔表面缺陷的实时检测。同时，算法具有较好的扩展性，能够方便地应用到其他产品的表面缺陷检测中。 

Robust Principal Component Analysis for Aluminum Foil Surface Defects Detection

In order to accurately detect the defects on the aluminum surface, including perforation, stains, shine marks and scratches, etc., a method of detection of aluminum foil images surface defects based on low-rank sparse decomposition is proposed. The probability of aluminum foil surface defects occurring in the production process is low and the defects typically constitute a small area proportion on a foil image. In other words, a linear relationship may be presumed between a foil image and the background, which may both be roughly deemed to under the same low-rank subspace, and a further presumption is that surface defects are approximately sparse. The observation data matrix consisting of a foil image sequence is put to a low-rank sparse decomposition using RPCA(Robust Principal Component Analysis) algorithm, resulting in low-rank background images and sparse defective images. A low-rank sparse decomposition test is performed on image sequences consisting respectively of single foil images and multiple foil images, and the validity of the proposed technique is demonstrated in an applicable detection process of foil image defects. Experimental results showed that the proposed algorithm detect defects were clear and complete, processing of an image with a size of 880×540 takes no more than 0.7 second on average, and it’s able to realize real-time detections. Algorithms presented in this paper are featured with rather favorable expansibility, and it can be easily applied into surface defect detections for other objects.