Anomaly Detection Algorithm of Ballastless Track Bed Based on Image Inpainting

Fig. 3. Masking process of the MASK module. (a) Input image; (b) mask map at $k_{j}$ ; (c) input image after mask; (d) reconstructed image; (e) complement mask map of the mask map at $k_{j}$ ; (f) partial reconstructed image after mask; (g) final reconstructed image

Download full size

Fig. 4. Public dataset presentation. (a) MNIST; (b) CIFAR-10

Download full size

Fig. 5. Real ballastless track bed dataset display 1. (a) Normal samples; (b) abnormal samples

Download full size

Fig. 6. Real ballastless track bed dataset display 2. (a) Normal samples; (b) abnormal samples

Download full size

Fig. 7. Comparison of reconstruction effects of different algorithms on MNIST dataset. (a) Original image; (b) AnoGAN; (c) GANomaly; (d) Skip-GANomaly; (e) proposed algorithm

Download full size

Fig. 8. Normal image reconstruction effect and comparison of abnormal images of different algorithms on ballastless track bed dataset. (a) Normal test image; (b) rendering diagram of AnoGAN algorithm; (c) rendering diagram of GANomaly algorithm; (d) rendering diagram of Skip-GANomaly algorithm; (e) rendering diagram of the proposed algorithm; (f) abnormal diagram of the picture tested by AnoGAN algorithm; (g) abnormal diagram of the picture tested by GANomaly algorithm; (h) abnormal diagram of the picture tested by Skip-GANomaly algorithm; (i) average anomaly map of the proposed algorithm in multiple scales; (j) anomaly map obtained from threshold determination using the proposed algorithm

Download full size

Fig. 9. Comparison of abnormal image reconstruction of different algorithms on ballastless track bed dataset

Download full size

Fig. 10. Comparison of abnormal images obtained from abnormal image reconstruction errors of different algorithms on ballastless track bed dataset

Download full size

Fig. 11. Visualization results of abnormal images processed by the proposed algorithm on ballastless track line dataset

Download full size

Table 1. AUC of different abnormality detection algorithms for each class on the public dataset MNIST

View table

Table 1. AUC of different abnormality detection algorithms for each class on the public dataset MNIST

Model	0	1	2	3	4	5	6	7	8	9
AnoGAN	0.610	0.300	0.535	0.440	0.430	0.420	0.475	0.335	0.400	0.335
GANomaly	0.850	0.330	0.920	0.771	0.770	0.800	0.880	0.780	0.890	0.580
Skip-GANomaly	0.826	0.998	0.701	0.863	0.640	0.574	0.785	0.763	0.765	0.720
Ours	0.978	0.999	0.821	0.857	0.953	0.901	0.945	0.998	0.978	0.901

Table 2. AUC of different abnormality detection algorithms for each class on the public dataset CIFAR-10

View table

Table 2. AUC of different abnormality detection algorithms for each class on the public dataset CIFAR-10

Model	dog	bird	cat	car	frog	deer	plane	ship	truck	horse
AnoGAN	0.393	0.411	0.399	0.492	0.321	0.335	0.516	0.567	0.511	0.399
GANomaly	0.660	0.562	0.612	0.644	0.923	0.778	0.954	0.861	0.719	0.638
Skip-GANomaly	0.624	0.594	0.582	0.769	0.937	0.816	0.985	0.892	0.739	0.654
Ours	0.825	0.745	0.891	0.921	0.967	0.837	0.986	0.957	0.967	0.876

Table 3. Average AUC of the different abnormality detection networks on public datasets MNIST and CIFAR-10
View table
Table 3. Average AUC of the different abnormality detection networks on public datasets MNIST and CIFAR-10
Model MNIST CIFAR-10
AnoGAN 0.472 0.434
GANomaly 0.715 0.735
Skip-GANomaly 0.787 0.760
Ours 0.933 0.897

Table 4. Performance of different algorithms on ballastless track bed dataset
View table
Table 4. Performance of different algorithms on ballastless track bed dataset
Model DR /% AR /% OR /% FAR /% AUC
AnoGAN 46.71 56.42 53.29 34.02 0.60
GANomaly 74.89 79.43 25.11 16.00 0.85
Skip-GANomaly 96.54 53.64 3.46 90.02 0.52
Ours 97.04 94.50 2.96 8.51 0.95

Table 5. General performance of the proposed algorithm
View table
Table 5. General performance of the proposed algorithm
Model DR /% AR /% OR /% FAR /% AUC
Ours 93.75 89.10 6.25 12.50 0.92

Table 6. Ablation experimental performance of the proposed algorithm on ballastless track bed dataset

View table

Table 6. Ablation experimental performance of the proposed algorithm on ballastless track bed dataset

Dataset	Model	DR /%	AR /%	OR /%	FAR /%
Ballastless track line 1	Ours（no）	74.89	79.43	25.11	16.00
	Ours（MASK）	96.10	77.45	3.90	41.50
	Ours（MASK-max）	97.04	94.50	2.96	8.51
Ballastless track line 2	Ours（no）	21.91	42.23	78.09	37.50
	Ours（MASK）	81.25	80.00	18.75	21.88
	Ours（MASK-max）	93.75	89.10	6.25	12.50

Tools

Get Citation

Copy Citation Text

Wan Jiang, Kai Yang, Chunrong Qiu, Liming Xie. Anomaly Detection Algorithm of Ballastless Track Bed Based on Image Inpainting[J]. Laser & Optoelectronics Progress, 2024, 61(12): 1237006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: May. 15, 2023

Accepted: Sep. 18, 2023

Published Online: May. 29, 2024

The Author Email: Kai Yang (yangkai_swjtu@163.com)

DOI:10.3788/LOP231318

CSTR:32186.14.LOP231318

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. AUC of different abnormality detection algorithms for each class on the public dataset MNIST

Table 1. AUC of different abnormality detection algorithms for each class on the public dataset MNIST

Table 2. AUC of different abnormality detection algorithms for each class on the public dataset CIFAR-10

Table 2. AUC of different abnormality detection algorithms for each class on the public dataset CIFAR-10

Table 3. Average AUC of the different abnormality detection networks on public datasets MNIST and CIFAR-10

Table 3. Average AUC of the different abnormality detection networks on public datasets MNIST and CIFAR-10

Table 4. Performance of different algorithms on ballastless track bed dataset

Table 4. Performance of different algorithms on ballastless track bed dataset

Table 5. General performance of the proposed algorithm

Table 5. General performance of the proposed algorithm

Table 6. Ablation experimental performance of the proposed algorithm on ballastless track bed dataset

Table 6. Ablation experimental performance of the proposed algorithm on ballastless track bed dataset