When Textures Deceive: Weakly Supervised Industrial Anomaly Detection with Adapted-Loss CycleGAN

Complex background textures challenge industrial anomaly detection (IAD) due to their intricate spatial variations, yet they are underrepresented in datasets like MVTecAD. Pixel-wise labeling is impractical, while existing weakly supervised and unsupervised methods largely overlook such textures due to the lack of suitable benchmarks. Moreover, subtle domain discrepancies in textures and occluded anomalies can degrade the performance of unsupervised approaches. GAN-based approaches, though widely used, have inherent loss function flaws reducing the effectiveness of IAD in complex environments. We introduce the Manufacturing Complex Background Texture (MCBT) dataset with 1,027 real-world images of diverse machining-induced textures, and propose Adapted-Loss Cycle-Consistent GAN (AL-CycleGAN) that leverages domain transfer while mitigating GAN limitations via a power-switch algorithm. Our analysis demonstrates that (1) MCBT introduces greater texture complexity while complementing standard datasets, (2) AL-CycleGAN achieves state-of-the-art performance on both MCBT and existing benchmarks, and (3) the proposed approach significantly improves GAN-based IAD by addressing critical loss function limitations.