Efficient textile anomaly detection via memory guided distillation network

Textile anomaly detection with high accuracy and fast frame rates are desired in real industrial scenarios. To this end, we propose an efficient memory guided distillation network, which includes encoder, decoder, and segmentation networks. Instead of utilizing a pre-trained large network as the encoder, we utilize a small feature extraction network, whose features are distilled from a teacher network. To improve the reconstruction quality with small networks, we further introduce an efficient memory bank, whose features are extracted by the teacher network with normal reference inputs. Considering the blurry reconstruction may lead to false-positive results, we further introduce a pseudo-normal simulation method by augmenting the inputs with blurry effects. Besides, we construct a Textile Anomaly dataset (Textile AD) for textile anomaly detection with pixel-wise labels for comprehensively evaluation and our method demonstrates superior performance on the Textile AD dataset. Additionally, we performed experiments using the publicly accessible MVTec-AD industrial anomaly dataset and our approach aligns closely with the performance of cutting-edge methodologies, which demonstrates that our method is applicable to other industrial product categories. Our Textile AD is shared in https://github.com/Songziyangtju/Textile-AD-dataset.