Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss Function
Monitoring the health of infrastructure is critical to maintaining the integrity of concrete construction. Conventional crack detection methods that rely on visual inspection and image processing often produce inconsistent results. U-Net, an architecture often used in image processing, has limitatio...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10855421/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1832540510339727360 |
|---|---|
| author | Andrew Prasetyo I Ketut Eddy Purnama Eko Mulyanto Yuniarno Priyo Suprobo |
| author_facet | Andrew Prasetyo I Ketut Eddy Purnama Eko Mulyanto Yuniarno Priyo Suprobo |
| author_sort | Andrew Prasetyo |
| collection | DOAJ |
| description | Monitoring the health of infrastructure is critical to maintaining the integrity of concrete construction. Conventional crack detection methods that rely on visual inspection and image processing often produce inconsistent results. U-Net, an architecture often used in image processing, has limitations in crack segmentation, especially regarding the loss function. The main challenge is class imbalance, as cracks usually occupy a much smaller area compared to the background in concrete images. To address this, we developed an automated technique utilizing the U-Net framework with a novel loss function that we named DBCE Loss. Unlike the typical BCED we added the LogCoshDice function to get an improvement so that the resulting loss will be better and can improve the performance of the model. The performance of our method is rigorously evaluated by combining DeepCrack, GAPS, Crack500, and CrackForest datasets to illustrate that the model can detect cracks in various material conditions. DeepCrack represents cracks in concrete with minimal distress, while GAPS represents cracks in asphalt. Crack500 covers cracks with significant distress such as gravel, and CrackForest focuses on small cracks in concrete.The proposed U-Net model achieved Maximum accuracies between 98.58% and 99.22%, Maximum Dice coefficients from 88.27% to 98.03%, and Maximum F1-scores up to 97.11%. The proposed method is 7.69% ahead of the largest comparison method (DICE Loss) on Average IOU, while on Average Precission is leading 6.72% ahead of the largest comparison method (DICE). Then in terms of Recall our method is 8.84% superior to the largest comparison method (BCE). Sera in terms of Average F1-Score is 14.64% ahead of the largest comparison method (DICE). These results show that our method has surpassed conventional loss methods such as BCE, DICE and BCED. This research can facilitate a more reliable infrastructure health monitoring process and help with crack detection in the future. |
| format | Article |
| id | doaj-art-18c8eb5a13304a1cae6ce95a84120b23 |
| institution | Kabale University |
| issn | 2169-3536 |
| language | English |
| publishDate | 2025-01-01 |
| publisher | IEEE |
| record_format | Article |
| series | IEEE Access |
| spelling | doaj-art-18c8eb5a13304a1cae6ce95a84120b232025-02-05T00:01:04ZengIEEEIEEE Access2169-35362025-01-0113209032092210.1109/ACCESS.2025.353480310855421Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss FunctionAndrew Prasetyo0https://orcid.org/0009-0001-6489-5304I Ketut Eddy Purnama1https://orcid.org/0000-0002-7438-7880Eko Mulyanto Yuniarno2https://orcid.org/0000-0003-1243-3025Priyo Suprobo3Department of Electrical Engineering, Institut Teknologi Sepuluh Nopember, Surabaya, IndonesiaDepartment of Electrical Engineering, Institut Teknologi Sepuluh Nopember, Surabaya, IndonesiaDepartment of Electrical Engineering, Institut Teknologi Sepuluh Nopember, Surabaya, IndonesiaDepartment of Civil Engineering, Institut Teknologi Sepuluh Nopember, Surabaya, IndonesiaMonitoring the health of infrastructure is critical to maintaining the integrity of concrete construction. Conventional crack detection methods that rely on visual inspection and image processing often produce inconsistent results. U-Net, an architecture often used in image processing, has limitations in crack segmentation, especially regarding the loss function. The main challenge is class imbalance, as cracks usually occupy a much smaller area compared to the background in concrete images. To address this, we developed an automated technique utilizing the U-Net framework with a novel loss function that we named DBCE Loss. Unlike the typical BCED we added the LogCoshDice function to get an improvement so that the resulting loss will be better and can improve the performance of the model. The performance of our method is rigorously evaluated by combining DeepCrack, GAPS, Crack500, and CrackForest datasets to illustrate that the model can detect cracks in various material conditions. DeepCrack represents cracks in concrete with minimal distress, while GAPS represents cracks in asphalt. Crack500 covers cracks with significant distress such as gravel, and CrackForest focuses on small cracks in concrete.The proposed U-Net model achieved Maximum accuracies between 98.58% and 99.22%, Maximum Dice coefficients from 88.27% to 98.03%, and Maximum F1-scores up to 97.11%. The proposed method is 7.69% ahead of the largest comparison method (DICE Loss) on Average IOU, while on Average Precission is leading 6.72% ahead of the largest comparison method (DICE). Then in terms of Recall our method is 8.84% superior to the largest comparison method (BCE). Sera in terms of Average F1-Score is 14.64% ahead of the largest comparison method (DICE). These results show that our method has surpassed conventional loss methods such as BCE, DICE and BCED. This research can facilitate a more reliable infrastructure health monitoring process and help with crack detection in the future.https://ieeexplore.ieee.org/document/10855421/Crack segmentationU-Netloss functiondeep learningstructural health monitoring |
| spellingShingle | Andrew Prasetyo I Ketut Eddy Purnama Eko Mulyanto Yuniarno Priyo Suprobo Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss Function IEEE Access Crack segmentation U-Net loss function deep learning structural health monitoring |
| title | Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss Function |
| title_full | Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss Function |
| title_fullStr | Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss Function |
| title_full_unstemmed | Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss Function |
| title_short | Improving Crack Detection Precision of Concrete Structures Using U-Net Architecture and Novel DBCE Loss Function |
| title_sort | improving crack detection precision of concrete structures using u net architecture and novel dbce loss function |
| topic | Crack segmentation U-Net loss function deep learning structural health monitoring |
| url | https://ieeexplore.ieee.org/document/10855421/ |
| work_keys_str_mv | AT andrewprasetyo improvingcrackdetectionprecisionofconcretestructuresusingunetarchitectureandnoveldbcelossfunction AT iketuteddypurnama improvingcrackdetectionprecisionofconcretestructuresusingunetarchitectureandnoveldbcelossfunction AT ekomulyantoyuniarno improvingcrackdetectionprecisionofconcretestructuresusingunetarchitectureandnoveldbcelossfunction AT priyosuprobo improvingcrackdetectionprecisionofconcretestructuresusingunetarchitectureandnoveldbcelossfunction |