AGSSP

Official PyTorch Implementation of Advancing Metallic Surface Defect Detection via Anomaly-Guided Pretraining on a Large Industrial Dataset

Pretrained models are commonly employed to improve finetuning performance in metallic surface defect detection, especially in data-scarce environments. However, models pretrained on ImageNet often underperform due to data distribution gaps and misaligned training objectives. To address this, we propose a novel method called Anomaly-Guided Self-Supervised Pretraining (AGSSP), which pretrains on a large industrial dataset containing 120,000 images. AGSSP adopts a two-stage framework: (1) anomaly map guided backbone pretraining, which integrates domain-specific knowledge into feature learning through anomaly maps, and (2) anomaly box guided detector pretraining, where pseudo-defect boxes derived from anomaly maps act as targets to guide detector training. Anomaly maps are generated using a knowledge enhanced anomaly detection method. Additionally, we present two small-scale, pixel-level labeled metallic surface defect datasets for validation. Extensive experiments demonstrate that AGSSP consistently enhances performance across various settings, achieving up to a 10% improvement in mAP@0.5 and 11.4% in mAP@0.5:0.95 compared to ImageNet-based models.

Environments

# Our code is based on mmyolo=0.6.0, mmpretrain=1.2.0, mmdet=3.3.0, mmseg=1.2.2.
pip3 install -r requirements.txt

Anomaly Map Generation

# generate anomaly map
cd KeAD/
sh zero_shot_plain_base.sh

# anomaly map to yolo txt
python generate_yolo_box_from_anomaly.py

Pre-training and Fine-tuning Instructions

Pre-training

# Change the corresponding configuration in configs
sh run_pretrain.sh

# multi gpu
sh dist_run_pretrain.sh

Fine-tuning

# Change the corresponding configuration in configs
sh run_yolo.sh

Results of different Methods

Backbone	Detector	Pretrain Method	Pretrain Dataset	mAP@0.5	mAP@0.5:0.95	Pretrain Weight	Config
CSPDarket	YOLO	cls	ImageNet	71.6	33.5	-	-
CSPDarket	YOLO	cls	Industrial	72.8	34.7	-	-
CSPDarket	YOLO	AGBP	Industrial	74.9	37.5	-	-
CSPDarket	YOLO	AGSSP	Industrial	77.2	44.9	baiduyun / google	config
ResNet50	YOLO	cls	ImageNet	69.1	32.3	-	config
ResNet50	YOLO	cls	Industrial	69.1	33.6	-	-
ResNet50	YOLO	AGBP	Industrial	72.0	36.1	-	-
ResNet50	YOLO	AGSSP	Industrial	74.5	42.5	baiduyun / google	config
Swin-Base	YOLO	cls	ImageNet	75.7	37.7	-	config
Swin-Base	YOLO	cls	Industrial	73.3	36.0	-	-
Swin-Base	YOLO	AGBP	Industrial	76.8	39.0	-	-
Swin-Base	YOLO	AGSSP	Industrial	77.7	45.0	baiduyun / google	config
Swin-Base	YOLO	SimMIM	ImageNet	78.3	44.3	-	-
Swin-Base	YOLO	SimMIM	Industrial	76.5	41.5	-	-
Swin-Base	YOLO	AGBP	Industrial	79.8	44.3	-	-
Swin-Base	YOLO	AGSSP	Industrial	78.8	48.4	baiduyun / google	config
Swin-Base	Faster R-CNN	SimMIM	ImageNet	64.6	31.7	-	config
Swin-Base	Faster R-CNN	AGBP	Industrial	71.1	38.6	-	config
Swin-Base	Faster R-CNN	AGSSP	Industrial	73.0	41.0	baiduyun / google	config
Swin-Base	DINO	SimMIM	ImageNet	78.6	49.3	-	config
Swin-Base	DINO	AGBP	Industrial	79.2	49.5	-	config
Swin-Base	DINO	AGSSP	Industrial	79.0	50.4	baidyun / google	config

Performance Comparison of Different Backbones and Pretraining Methods

Citation

If you find our work is useful in your research or applications, please consider giving us a star 🌟 and citing it.

@misc{liu2025advancingmetallicsurfacedefect,
  title={Advancing Metallic Surface Defect Detection via Anomaly-Guided Pretraining on a Large Industrial Dataset},
  author={Chuni Liu and Hongjie Li and Jiaqi Du and Yangyang Hou and Qian Sun and Lei Jin and Ke Xu},
  year={2025},
  eprint={2509.18919},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2509.18919},
}