Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Research on Weakly Supervised Pavement Crack Segmentation Based on Defect Location by Generative Adversarial Network and Target Re‐optimization
https://orcid.org/0000-0002-1049-4599
https://orcid.org/0009-0009-8248-3483
Abstract This paper presents a weakly supervised method of pavement crack segmentation. The U-GAT-IT model is trained by image-level labels to obtain the class activation map (CAM). Then, by erasing the cracks that have been detected, the image is re-detected to form a more complete CAM. The pseudo-labels obtained by using the adaptive threshold segmentation method are used to train U2-Net, and the trained U2-Net is used to optimize the pseudo-labels to train U2-Net with better quality. Finally, the result shows that the method effectively narrows the gap between unsupervision and full supervision of pavement cracks.
Highlights A method of weakly supervised pavement crack segmentation is proposed, which can effectively narrow the performance gap between weakly supervised and fully supervised methods. The location of pavement cracks is obtained by CAM based on a generative adversarial network (U-GAT-IT), and a secondary optimization method of CAM to completely cover the target area is proposed. An adaptive threshold segmentation method based on CAM images is proposed to separate cracks with more details. A research idea for obtaining better pseudo-label quality is put forward that uses pseudo-labels to train the segmentation network and optimizes them with the obtained model.
Research on Weakly Supervised Pavement Crack Segmentation Based on Defect Location by Generative Adversarial Network and Target Re‐optimization
https://orcid.org/0000-0002-1049-4599
https://orcid.org/0009-0009-8248-3483
Abstract This paper presents a weakly supervised method of pavement crack segmentation. The U-GAT-IT model is trained by image-level labels to obtain the class activation map (CAM). Then, by erasing the cracks that have been detected, the image is re-detected to form a more complete CAM. The pseudo-labels obtained by using the adaptive threshold segmentation method are used to train U2-Net, and the trained U2-Net is used to optimize the pseudo-labels to train U2-Net with better quality. Finally, the result shows that the method effectively narrows the gap between unsupervision and full supervision of pavement cracks.
Highlights A method of weakly supervised pavement crack segmentation is proposed, which can effectively narrow the performance gap between weakly supervised and fully supervised methods. The location of pavement cracks is obtained by CAM based on a generative adversarial network (U-GAT-IT), and a secondary optimization method of CAM to completely cover the target area is proposed. An adaptive threshold segmentation method based on CAM images is proposed to separate cracks with more details. A research idea for obtaining better pseudo-label quality is put forward that uses pseudo-labels to train the segmentation network and optimizes them with the obtained model.
Research on Weakly Supervised Pavement Crack Segmentation Based on Defect Location by Generative Adversarial Network and Target Re‐optimization
https://orcid.org/0000-0002-1049-4599
https://orcid.org/0009-0009-8248-3483
Abstract This paper presents a weakly supervised method of pavement crack segmentation. The U-GAT-IT model is trained by image-level labels to obtain the class activation map (CAM). Then, by erasing the cracks that have been detected, the image is re-detected to form a more complete CAM. The pseudo-labels obtained by using the adaptive threshold segmentation method are used to train U2-Net, and the trained U2-Net is used to optimize the pseudo-labels to train U2-Net with better quality. Finally, the result shows that the method effectively narrows the gap between unsupervision and full supervision of pavement cracks.
Highlights A method of weakly supervised pavement crack segmentation is proposed, which can effectively narrow the performance gap between weakly supervised and fully supervised methods. The location of pavement cracks is obtained by CAM based on a generative adversarial network (U-GAT-IT), and a secondary optimization method of CAM to completely cover the target area is proposed. An adaptive threshold segmentation method based on CAM images is proposed to separate cracks with more details. A research idea for obtaining better pseudo-label quality is put forward that uses pseudo-labels to train the segmentation network and optimizes them with the obtained model.
Research on Weakly Supervised Pavement Crack Segmentation Based on Defect Location by Generative Adversarial Network and Target Re‐optimization
He, Tiejun (Autor:in) / Li, Huaen (Autor:in) / Qian, Zhendong (Autor:in) / Niu, Chenyi (Autor:in) / Huang, Ruihua (Autor:in)
16.12.2023
Aufsatz (Zeitschrift)
Elektronische Ressource
Englisch
| Wiley | 2025