Rethinking unsupervised texture defect detection using PCA

Abstract

In recent times, deep learning frameworks have become the preferred solution for visual object detection tasks, including texture defect detection. However, in real-world industrial applications, such as operation and maintenance of gas-insulated switchgear (GIS) equipment, collecting a large number of samples in advance is difficult because opening GIS equipment affects its normal operation; moreover, only four real-world GIS equipment defect samples are available currently. Most deep learning methods perform poorly when trained using a small number of labelled samples or when there is a lack of computational resources. In addition, the features extracted by multi-layer convolution and down-sampling are often unsuitable for small-size and low-contrast defects. However, the structure of the repetitive textured surface is simple and easy to model. Hence, this paper proposes a texture-defect detection method using principal components analysis (PCA) and histogram-based outlier score (HBOS) that requires only a small number of unlabelled samples and low computational complexity. First, PCA is utilized to extract features from the training patches to enhance the difference between the defect and background regions. Then, non-saliency suppression is applied for feature fusion to obtain the saliency map. Finally, the saliency value is converted to the defect score, and the defect is located by HBOS. Experiments on two public datasets and a GIS simulation equipment demonstrate that the proposed method is suitable for product surface defect detection and operation and maintenance in power distribution equipment. Moreover, compared with traditional texture feature-based and deep learning methods, the proposed method yields better results for small-size and low-contrast defects with only four unlabelled training samples per class.