SCL-VI: Self-supervised Context Learning for Visual Inspection of Industrial Defects

We address the challenge of detecting object defects through the self-supervised learning approach of solving the jigsaw puzzle problem.

Results

Dependencies

Since I did this project a long time ago, there may be some potential issues with environmental dependencies.

• Tested with Python 3.8
• Pytorch v1.6.0

Dateset

• Dataset : MvTec AD

Run Training

• python train.py --obj=cable --lambda_value=1 --D=64 --epoches=400 --lr=1e-4 --gpu=0

Run Affinity Testing

• python test.py --obj=cable --gpu=0
• enc.load(obj, N) N is the serial number of the obtained training weight file

Anomaly maps

• python heat_map.py --obj=cable
• enc.load(obj, N) N is the serial number of the obtained training weight file

Details:

• The input of the network should be 256x256
• data.npy contains the relative positions and their reference numbers.