Manhattan Correlation Attention Network for Metal Part Anomaly Classification

Abnormal classification of metal components plays an important role in industrial product manufacturing. However, it is difficult to detect metal anomalies due to the following issues: (1) Some defects of metal parts areas are small; (2) The abnormal metal regions are relatively similar to the normal ones. To address these problems, a Manhattan Correlation Attention Network (MCA-Net) is proposed for anomaly classification from the metal part, where the Manhattan Retentive Attention (MRA) module is designed to search for small anomaly regions by global information modeling, and the Structural Contextual Attention (SCA) module is devised to discriminate the similar abnormal regions from the normal ones by aggregating contextual structured dependency. Experiments on the benchmark verify the effectiveness of the proposed MCA-Net for metal part anomaly classification, achieving the performance of 91.76%, 91.76%, 91.69%, 91.35% on the accuracy, precision, specificity, and F1-score, respectively, further assisting in metal part classification in manufacturing.