Official Emphasizing Differences repository

we propose a novel anomaly detection framework called Emphasize Differences (ED), which performs feature reconstruction by jointly learning intra-image and inter-image differences. Inspired by human visual discrepancy detection, the ED framework explicitly amplifies these two types of differential signals during feature reconstruction. It achieves this through dual-path encoding: intra-image differences are captured via self-supervised constraints to identify local anomalies that disrupt structural consistency, while inter-image differences are modeled by a contrastive memory module that aligns reconstructed features with normal prototypes, thereby isolating anomalous responses in latent space.

In this work, we propose a novel anomaly detection framework called Emphasize Differences (ED), which addresses the "learning shortcut" problem in reconstruction-based methods by simultaneously modeling feature discrepancies within individual samples and variations across different samples. The ED framework begins by extracting multi-scale feature maps from input images using a CNN backbone pre-trained on ImageNet. These reference features, combined with the training data, are then processed by a Transformer network. The 2D input images are first transformed into a high-dimensional feature space via an Embedding2D module. Next, the features are passed through a deep network consisting of Difference-Aware Attention and D2Fusion blocks, which progressively learn both intra-sample relationships and inter-sample feature correlations.

• tqdm == 4.66.6
• numpy == 1.24.4
• matplotlib == 3.5.0
• scipy == 1.9.3
• scikit-learn == 1.1.3
• pandas == 2.0.3
• pillow == 9.1.0
• torch == 1.13.1
• torchvision == 0.14.1
• timm == 0.6.12
• huggingface-hub == 0.26.2

python create_distance_maps.py 

#  MVTecAD
python create_ref_features.py --dataset mvtec --data_path xxx --backbone_arch xxx --save_path xxx
#  BTAD
python create_ref_features.py --dataset btad --data_path xxx --backbone_arch xxx --save_path xxx
#  MVTec3D-RGB
python create_ref_features.py --dataset mvtec3d --data_path xxx --backbone_arch xxx --save_path xxx

python main.py --dataset mvtec --data_path /path/to/your/dataset --backbone_arch xxx --rfeatures_path xxx --with_intra --with_inter --save_prefix xxx

python main.py --dataset btad --data_path /path/to/your/dataset --backbone_arch xxx --rfeatures_path xxx --with_intra --with_inter --save_prefix xxx

python main.py --dataset mvtec3d --data_path /path/to/your/dataset --backbone_arch xxx --rfeatures_path xxx --with_intra --with_inter --save_prefix xxx

Thanks for the great efforts from FOD.