MemCLR-GANomaly: An Unsupervised Anomaly Detection Approach for Resin Cut-Off Wheel Surfaces

Unsupervised anomaly detection is critical for industrial surface inspection, especially for resin cut-off wheels where defect samples are scarce and morphologically diverse. GANomaly serves as a classic reconstruction-based baseline using latent vector discrepancy, yet its single-score mechanism conflates pixel-level structural, feature-level semantic, and distribution-level statistical anomalies into one scalar, causing inherent detection blind spots for complex industrial defects. To address this issue, we propose a three-dimensional anomaly scoring framework instantiated as MemCLR-GANomaly. Specifically, a multi-scale region error module captures pixel-level structural anomalies; a SimCLR-enhanced contrastive discrepancy branch quantifies feature-level semantic deviations; and a dynamic memory bank models distribution-level statistical normality. An adaptive fusion network integrates the three indicators with sample-dependent weights to generate the final anomaly score. Experiments on the self-built Resin Cut-off Wheel Dataset achieve image- and pixel-level AUROC values of 0.995 and 0.988, with a false positive rate as low as 0.012. On a custom partition of the public MVTec AD dataset, the method attains an image-level AUROC of 0.948 and a pixel-level AUROC of 0.936. With an inference speed of 42.3 FPS and a computational cost of 3.9 GFLOPs, the proposed method meets the real-time requirements of online industrial inspection for resin cut-off wheel production lines.