MaMiNet: Memory-attended multi-inference network for surface-defect detection

Surface-defect detection has attracted extensive attention in the field of industrial inspection but remains challenging, owing to the rare occurrence and the various appearance of the defects. Promising results have been obtained by supervised methods but they require a large number of pixel-level annotations which are very costly to obtain. This paper proposes a memory-attended multi-inference network (MaMiNet) for image-level defect detection. MaMiNet integrates image classification with saliency detection and can accommodate a variable number of samples with pixel-label annotations along with image-level annotation. Considering the various defect appearance, a memory attention feature enhancement module is exploited to capture the attention information not only within one sample but across whole samples and seek better representation ability for the defective regions. A multi-inference aware aggregation module is proposed to fuse features with different inference hints and obtain more comprehensive features. The proposed method is extensively validated on four datasets and better experimental results are obtained compared with other state-of-the-art methods, especially in weak supervision mode without any pixel-level annotation. The efficacy of the proposed modules is validated through ablation studies.