Scholarlink

Monitoring damage in concrete structures is essential for maintaining their integrity. Utilizing computer vision is crucial for achieving precise and quantitative monitoring. Recent advancements have integrated computer vision with deep learning, particularly convolutional neural network (CNN) models. CNNs are known for their accuracy and adaptability in various contexts. They have been designed for pixel-by-pixel image classification, specifically for pix-el-level detection, which is particularly effective for identifying and categoriz-ing damage with high detail. In practical scenarios, environments are often complex with many non-concrete elements. Therefore, this study aims to use pixel-level CNNs to detect damage on concrete surfaces. The research uses DeepLabV3 and DeepLabV3+ architectures for concrete crack segmentation. These models are evaluated and compared based on specific metrics and predic-tion outcomes. The results indicate that DeepLabV3 slightly outperforms DeepLabV3+ with Intersection over Union (IoU) values of 0.936 and 0.924 on the testing set, respectively. However, DeepLabV3 requires significantly more training time than DeepLabV3+. These findings demonstrate the practicality and effectiveness of these models for segmenting concrete cracks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Segmentation of Concrete Surface Cracks Using DeeplabV3 and DeeplabV3+

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Luong V. Le,
  • Nhi V. Nguyen,
  • Thong N. Truong,
  • Thanh Danh Tran

摘要

Monitoring damage in concrete structures is essential for maintaining their integrity. Utilizing computer vision is crucial for achieving precise and quantitative monitoring. Recent advancements have integrated computer vision with deep learning, particularly convolutional neural network (CNN) models. CNNs are known for their accuracy and adaptability in various contexts. They have been designed for pixel-by-pixel image classification, specifically for pix-el-level detection, which is particularly effective for identifying and categoriz-ing damage with high detail. In practical scenarios, environments are often complex with many non-concrete elements. Therefore, this study aims to use pixel-level CNNs to detect damage on concrete surfaces. The research uses DeepLabV3 and DeepLabV3+ architectures for concrete crack segmentation. These models are evaluated and compared based on specific metrics and predic-tion outcomes. The results indicate that DeepLabV3 slightly outperforms DeepLabV3+ with Intersection over Union (IoU) values of 0.936 and 0.924 on the testing set, respectively. However, DeepLabV3 requires significantly more training time than DeepLabV3+. These findings demonstrate the practicality and effectiveness of these models for segmenting concrete cracks.