yuride0404127/aclnet-kd

Asymmetrical Contrastive Learning Network via Knowledge Distillation for No-Service Rail Surface Defect Detection

ACLNet-KD

Asymmetrical Contrastive Learning Network via Knowledge Distillation for No-Service Rail Surface Defect Detection

The model structure is as follows：

Requirements

Python >=3.6

Pytorch >= 1.7.1

Cuda = 10.2

Feature Maps

The results graphs of our tests are saved in the：

[Baidu Clould]https://pan.baidu.com/s/1d0poeuS5JQt4dDQ9FhibJg?pwd=cim9 提取码：cim9

Training Weights

All the training weights of our model will be saved in the:

[Baidu Clould] https://pan.baidu.com/s/13FeF2hgFtm6SrsEkhnT-lQ?pwd=dsy6 提取码：dsy6

Comparison of results table

Table I Evaluation metrics obtained from compared methods.

Model	Sm ↑	MAE ↓	maxFm ↑	maxEm ↑	FLOPs/G ↓	Params/M ↓
S2MA2020	0.775	0.141	0.817	0.864	108.1	86.7
EDR2020	0.811	0.082	0.850	0.893	32.2	39.3
BBS2020	0.828	0.074	0.867	0.909	12.7	49.8
HAI2021	0.718	0.171	0.803	0.829	73.5	59.8
EMI2021	0.800	0.104	0.850	0.876	106.9	99.1
SPNet2021	0.830	0.072	0.877	0.915	55.0	175.3
DAC2022	0.824	0.071	0.875	0.911	109.3	98.4
CSEP2022	0.814	0.085	0.866	0.899	45.4	18.8
CIR2022	0.809	0.086	0.856	0.894	17.3	103.2
RD3D2022	0.797	0.093	0.839	0.883	17.6	28.9
DRER2022	0.844	0.059	0.891	0.933	17.3	69.8
CLA2022	0.835	0.069	0.878	0.920	82.6	184.2
LENO2022	0.817	0.083	0.857	0.900	162.1	131.0
ICON2023	0.843	0.066	0.884	0.925	34.8	65.7
CAVER2023	0.809	0.090	0.848	0.891	63.0	93.8
ACLNet-T	0.860	0.055	0.897	0.938	20.7	79.1
ACLNet-S	0.842	0.063	0.883	0.926	5.6	11.9
ACLNet-S*	0.856	0.056	0.896	0.936	5.6	11.9

Table II. Test and distillation results on RGBD-SOD public datasets.

Model		NJU2K				NLPR				STERE
	Sm ↑	MAE ↓	maxFm ↑	maxEm ↑	Sm ↑	MAE ↓	maxFm ↑	maxEm ↑	Sm ↑	MAE ↓	maxFm ↑	maxEm ↑
S2MA2020	0.894	0.053	0.930	0.889	0.915	0.030	0.953	0.902	0.890	0.051	0.932	0.882
EDR2020	0.839	0.076	0.883	0.823	0.886	0.038	0.928	0.860	0.851	0.065	0.899	0.831
BBS2020	0.921	0.035	0.949	0.920	0.930	0.023	0.961	0.918	0.908	0.041	0.942	0.903
HAI2021	0.912	0.038	0.944	0.915	0.921	0.024	0.960	0.915	0.907	0.040	0.944	0.906
EMI2021	0.881	0.050	0.924	0.876	0.916	0.026	0.953	0.902	0.897	0.042	0.938	0.894
DAC2022	0.890	0.046	0.929	0.887	0.913	0.025	0.949	0.897	0.899	0.043	0.936	0.892
MC2022	0.902	0.042	0.939	0.900	0.918	0.025	0.956	0.907	0.903	0.042	0.945	0.898
CIR2022	0.925	0.035	0.955	0.927	0.933	0.023	0.966	0.924	0.917	0.039	0.950	0.916
DRER2022	0.906	0.038	0.943	0.907	0.915	0.024	0.953	0.901	0.895	0.042	0.943	0.891
LENO2022	0.838	0.073	0.888	0.824	0.878	0.040	0.921	0.845	0.856	0.062	0.906	0.840
ICON2023	0.893	0.051	0.937	0.891	0.904	0.032	0.951	0.885	0.899	0.047	0.945	0.890
CAVER2023	0.926	0.030	0.958	0.928	0.934	0.021	0.970	0.928	0.918	0.033	0.955	0.916
ACLNet-T	0.929	0.028	0.960	0.932	0.937	0.018	0.971	0.929	0.918	0.032	0.955	0.915
ACLNet-S	0.911	0.038	0.949	0.910	0.927	0.023	0.965	0.918	0.916	0.036	0.952	0.909
ACLNet-S*	0.913	0.037	0.951	0.914	0.930	0.021	0.966	0.919	0.917	0.035	0.953	0.911