ProtoDC-Net

Dataset download

1. USB-MD: https://drive.google.com/drive/folders/1NvQ5vZvZMdpJN8s1ttp9ZaMZ13OgQbAa?usp=sharing

2. USB-SD: https://github.com/Xavierman/A-deep-learning-based-surface-defect-inspection-system-using-multi-scale-and-channel-compressed-feat?tab=readme-ov-file

3. DAGM2007: https://conferences.mpi-inf.mpg.de/dagm/2007/prizes.html

Classification Accuracy

📊 Classification Accuracies on USB-MD Dataset

Method	BrightLine	Deformation	Dent	Scratch	Normal	Total
ETE	68.47	52.08	77.78	74.49	73.94	71.16
DECAF+MLR	50.45	52.60	92.07	78.59	80.03	73.38
SN_MRF_CC	75.23	69.27	96.83	80.44	78.51	79.02
FOMI	72.52	66.67	97.62	78.40	81.73	78.90
MHAF	74.33	85.94	97.62	77.47	85.96	82.55
ProtoDC-Net (ours)	85.14	84.90	96.83	79.14	90.69	86.03

Note: Bold values indicate the highest score per class.
All results are averaged over three runs with different random seeds (42, 43, and 44).

📊 Classification Accuracies on USB-SD Dataset

Method	BrightLine	Deformation	Dent	Scratch	Spot	Squalidity	Normal	Total
ETE	88.00	78.67	77.78	80.68	70.00	79.89	92.44	81.07
DECAF+MLR	91.89	88.00	87.33	79.89	66.22	83.11	96.33	84.68
SN_MRF_CC	94.56	91.67	96.34	93.78	91.33	94.66	97.44	94.25
SN_MRF_CC *	97.89	97.67	98.67	96.67	92.00	95.22	98.44	96.65
ProtoDC-Net	100.00	100.00	100.00	99.72	99.72	95.28	99.65	99.20

Note: * indicates that input images were augmented using rotation and cropping.
All results are averaged over three trials.

📊 Classification Accuracies on DAGM2007 Dataset

Method	Class 1	Class 2	Class 3	Class 4	Class 5	Class 6	Total
DECAF+MLR	36.67	83.33	5.67	80.00	6.67	70.00	47.06
SN_MRF_CC	100.00	100.00	93.33	100.00	30.00	100.00	87.22
ProtoDC-Net	100.00	100.00	100.00	100.00	93.33	100.00	98.89

Note: Classification accuracies (%) on six classes (Class 1–6) of the DAGM2007 dataset.
Each value indicates the classification accuracy on the defect class within each binary classification task.
All results are averaged over three trials.