Fast and Accurate Road Crack Detection Based on Adaptive Cost-Sensitive Loss Function.

Area of Science:

• Computer Vision
• Image Analysis
• Machine Learning

Background:

• Road crack detection is crucial for infrastructure maintenance.
• Significant foreground-background imbalance challenges existing detection methods.
• Loss function modification is a promising approach to address this imbalance.

Purpose of the Study:

• To propose a pixel-based adaptive weighted cross-entropy (WCE) loss function for high-quality road crack detection.
• To demonstrate the impact of loss functions on detection performance.
• To improve the efficiency of the crack detection training process.

Main Methods:

• Developed a pixel-based adaptive weighted cross-entropy (WCE) loss function.
• Integrated the proposed WCE loss with Jaccard distance for pixel-level detection.
• Conducted extensive experiments on four public datasets: CrackForest, AigleRN, Crack360, and BJN260.

Main Results:

• The proposed WCE loss significantly accelerates the training process compared to the vanilla WCE.
• The novel loss function effectively maintains high performance in road crack detection.
• Experimental results validate the superiority of the proposed method on multiple benchmarks.

Conclusions:

• Loss function design profoundly influences computer vision detection outcomes.
• The proposed adaptive WCE loss offers a sophisticated improvement for crack detection.
• This work provides valuable insights for future research in crack detection and similar imbalanced datasets.