Multi-stage unsupervised fabric defect detection based on DCGAN

The diversity of fabric defects and the lack of defect samples make detecting fabric defects an important and challenging problem. Currently, unsupervised algorithms are widely used for surface defect detection as they do not require annotated data and therefore reduce the cost of data acquisition. This paper presents a multi-stage unsupervised fabric defect detection method based on DCGAN. The method consists of three stages: the GAN training, the encoder training, and the classifier training. The first two stages allow our model to reconstruct the test images. In the image reconstruction process, we use a linear weighted fusion method to reduce the interference of defects. When the reconstructed image is subtracted from the original, we get a residual map that highlights the defects. This pixel-level detection makes it easier to detect different types of defects. In addition, we introduce a classifier training phase to generate a likelihood map for the test images. Each pixel value in the likelihood map represents the probability of the original map having a defect in that location region. Finally, we fuse the residual map with the likelihood map and further perform threshold segmentation on the fused residual map. Our method uses a separate training strategy at each stage and learns from a set of image patches cropped out online. The experimental results are compared with other methods in recent years and validate the method's superiority in terms of f-score metrics.

Automated summary

This paper proposes a multi-stage unsupervised fabric defect detection method based on DCGAN. Through image reconstruction and pixel-level detection, different types of defects can be accurately detected. In addition, the introduction of a classifier training phase and likelihood map further improves the accuracy of defect detection.