Fabric defect detection via a spatial cloze strategy

Deep-learning models have achieved state-of-the-art performances in a wide range of defect detection tasks. However, an inescapable criticism of one-stage fully supervised models is the lack of interpretability, which not only reduces the reliability of fabric defect detection systems but also limits the scope of their applications in production environments. To tackle the data imbalance and low interpretability of defect samples, we proposed a spatial cloze strategy for fabric defect detection, which reconstructs a local normal image and then feeds it into the detection model with the original image simultaneously. Specifically, we formulate the defect detection task as a novel image completion problem. Firstly, an end-to-end deep neural network is trained to finely restore the defect image by completing each image slice removed in sequence. Next, the progressive attention mechanism fuses the repaired normal image with the raw image, replacing the input layer of the cascade region-based convolutional neural network. Eventually, accurate instance-level defect segmentation can be obtained by comparing the repaired defect-free and the raw images. On the Tianchi dataset, the proposed method displays superior accuracy in 92% of defect classes, with a breakthrough in various categories that have hardly ever been detected. Extensive experiments on various complex fabric defect samples demonstrate that our strategy outperforms existing advanced methods.