Deep Metric Learning-Based Classification for Pavement Distress Images.

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Conventional pavement distress classification methods struggle with large dataset requirements and limited ability to learn new categories.
• High intra-class variance and low inter-class distinction in pavement distress images pose significant challenges for accurate automated analysis.

Purpose of the Study:

• To develop a novel deep metric learning-based method for pavement distress classification.
• To overcome limitations of existing methods, including data dependency and incremental learning.
• To enhance the accuracy and efficiency of road inspection through improved distress identification.

Main Methods:

• A Convolutional Neural Network (CNN) head with multi-cluster centroids trained using SoftTriple loss was designed.
• An adaptive weighting strategy was employed to address data imbalance by combining sample similarity and class priors.
• Soft-label techniques were utilized to reduce labeling noise by assessing similarity against support-set exemplars.

Main Results:

• The proposed method achieved superior performance on the UAV-PDD2023 dataset compared to traditional supervised learning.
• Demonstrated a 3.2% higher macro-recall than supervised learning methods.
• Showcased 6.7% and 8.5% improvements in macro-F1 and weighted-F1 scores, respectively, over iCaRL incremental learning.

Conclusions:

• The developed deep metric learning approach effectively classifies pavement distress, even with high intra-class variance and low inter-class distinction.
• The method's ability to incrementally learn new categories and handle data imbalance makes it suitable for real-world road inspection.
• This research offers a robust solution for evolving pavement distress types and scenarios with limited annotations, advancing automated road maintenance.