Reproduce “Spatial Attention Supervision for Defect Localization: Exploiting Ground-Truth Masks as Training Signal in Diffusion-Augmented Defect Detection” on OpenResearch

About this paper

Sajjad Rezvani Boroujeni, Muskan Saraf, Gnana Tulasi Makineni, Tom Bush, Hossein Abedi

Ground-truth defect masks are introduced as an auxiliary loss for image classifiers, encouraging their internal activation maps to align with defects without adding a segmentation head. On MVTec bottle inspection, the method improved activation-based Pixel-AUROC in four of eight CNN comparisons, including an 18.0% gain for EfficientNetB0 with DDPM augmentation, while leaving classification performance largely unchanged.

Key insights

• Spatial mask supervision can improve where a classifier looks without materially changing its image-level F1 score; the effect is strongest for models whose baseline activation maps are weak or misaligned.
• For EfficientNetB0, combining DDPM augmentation with attention guidance produced a significant interaction, raising activation-based Pixel-AUROC from 0.552 for the real-data standard baseline to 0.627.
• The method is architecture- and layer-dependent: ConvNeXt-T showed inconsistent results, transformers were not adapted to the loss, and Grad-CAM improvements were limited mainly to ResNet50.

Summary by alphaXiv.