High Power Laser Science and Engineering, Volume. 12, Issue 1, 010000e4(2024)
Continuous gradient fusion class activation mapping: segmentation of laser-induced damage on large-aperture optics in dark-field images
Figures & Tables(18)
Fig. 1. Schematic diagram of the methodology for online capturing images of optics (FODI images) by the FODI system.
Fig. 2. An example of the FODI image. (a)–(c) Images of stray light interference. (d), (e) Images of large damage sites. (f), (g) Images of weak damage sites.
Fig. 4. Examples of class activation maps generated by Grad-CAM and LayerCAM on FODI images. The red arrows point to scattered activated regions. The yellow arrows point to underactivated regions.
Fig. 6. The class activation maps of LayerCAM from different stages. The red box shows the feature and gradient maps of some channels from Stage 5.
Fig. 7. The original and CG-CAM methods of backpropagating gradients from the max-pooling layer to the convolution layer.
Fig. 8. Comparison of LayerCAM and CG-CAM results from different stages. (a), (d) Feature maps. (b), (e) Class activation gradient maps. (c), (f) Channel class activation maps.
Fig. 10. Comparison of the multiscale fusion effect from different stages of CG-CAM. The red arrows point to blurred boundaries.
Fig. 11. Examples of typical samples. (a) Background class samples. (b) Damage class samples. (c) Manually produced damage class samples.
Fig. 12. Comparison of the class activation maps and segmentation results between the baselines and our method. The green areas are the false positive segmentation results. The red areas are the segmentation results containing true damage sites.
Fig. 13. The overall damage segmentation results of a large-aperture optic. (a)–(d) Enlarged local images.
Fig. 14. Comparison of the class activation maps between LayerCAM and CG-CAM from different stages.
Fig. 15. The effect of each step in the nonlinear multiscale fusion algorithm.
Table 1. The classification performance of the VGG-16 model on our dataset.
View tableView in ArticleTable 1. The classification performance of the VGG-16 model on our dataset.
Model Accuracy Precision Recall FPR F1 VGG-16 97.54% 97.13% 98.39% 3.49% 97.75%
Table 2. Comparison of baselines and our method under various evaluation metrics.
View tableView in ArticleTable 2. Comparison of baselines and our method under various evaluation metrics.
Methods p-P (%) p-R (%) p-F1 (%) FDR (%) IoU (%) LASNR 63.39 99.27 70.18 41.49 37.21 VGG16-Unet 76.34 86.53 81.11 3.05 63.87 DeepLabv3 81.34 86.50 83.84 3.20 68.32 Grad-CAM 7.12 97.95 12.78 5.51 7.10 LayerCAM 61.90 89.73 70.89 19.57 41.82 Ours 84.24 93.55 87.32 0.90 63.78
Table 3. Comparison of the baseline and our two core algorithms under various evaluation metrics.
View tableView in ArticleTable 3. Comparison of the baseline and our two core algorithms under various evaluation metrics.
Methods p-P (%) p-R (%) p-F1 (%) FDR (%) IoU (%) LayerCAM 61.90 89.73 70.89 19.57 41.82 CG-CAM 75.52 94.87 82.53 8.68 52.17 CG-CAM+ NM-Fusion 84.24 93.55 87.32 0.90 63.78
Tools
Get Citation
Copy Citation Text
Yueyue Han, Yingyan Huang, Hangcheng Dong, Fengdong Chen, Fa Zeng, Zhitao Peng, Qihua Zhu, Guodong Liu. Continuous gradient fusion class activation mapping: segmentation of laser-induced damage on large-aperture optics in dark-field images[J]. High Power Laser Science and Engineering, 2024, 12(1): 010000e4
Paper Information
Category: Research Articles
Received: Jul. 27, 2023
Accepted: Nov. 7, 2023
Published Online: Feb. 2, 2024
The Author Email: Fengdong Chen (chenfd@hit.edu.cn), Guodong Liu (lgd@hit.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20