Automated fabric defect detection using multi-scale fusion MemAE

Fabric defect detection (FDD) is an essential part of the textile industry. To replace manual visual inspection, computer-vision-based solutions has been widely investigated in both unsupervised or supervised manner during the past few decades. However, precisely locating defective regions, especially for those tiny and inconspicuous ones, is still challenging. To this end, we propose an unsupervised learning approach for FDD via multi-scale memory-augmented auto-encoder (MemAE) fusion. The architecture of the proposed CNN model follows an encode-decode style. On the one hand, we explore the defect-free region reconstruction ability of the shallow CNN layers. On the other hand, the reconstruction ability of defect region is much emphasized in the deep layers. In the training phase of the network, we only use defect-free samples to learn the visual features of normal fabric textural representation. In the testing phase, a defective image is expected to be unpainted during the inference, and we further emphasize the defect regions by computing the pixel-level residual maps between the defective image and its corresponding reconstructed version. Additionally, a coarse-to-fine scheme is proposed to precisely locate the defect region by first introducing KNN and DeepSVDD to coarsely localize the defects, followed by a finer refinement using the obtained residual map. The proposed method is evaluated on two different types of datasets: the Periodic-pattern fabric database and the Yarn-dyed fabric database. The experimental results illustrate that our method consistently outperforms the state-of-the-art methods in terms of quality and robustness in both types of datasets with certain margin.