A Novel Approach to Pavement Crack Classification using Joint-Embedding Predictive Architectures

Abstract

One of the important issues in automating infrastructure maintenance is accurate and timely pavement distress detection. In this study, an efficient pavement distress classification model is proposed based on the hypothesis that self-supervised pre-training with the Joint Embedded Predictive Architecture (JEPA) can learn more abstract features as opposed to standard YOLOv11 methods. The proposed model specifically investigates the effectiveness of Global Average Pooling (GAP) in comparison with Class Token(CLS) as the standard approach. The proposed model is tested on standard dataset with high resolutions as well as in the target domain. The experimental evaluations on standard dataset showed the effectiveness of the proposed architecture of JEPA(GAP) with an F1-score of 99.50%, which is considerably better than the standard method in the detection of complex alligator cracks. The proposed model’s adaptability is also examined through experimental evaluations involving real-world datasets of 640 weather camera images with large domain gaps. By utilizing a Partial Freezing Strategy in JEPA, the model has shown a vast improvement in terms of F1-score of 73.68% on alligator cracks compared to YOLOv11’s 50.00% with an improvement of over 23%. The results validate that JEPA self-supervised learning coupled with effective aggregation of spatial features is far more effective in forming a sound base for generalization and adaptability in noisy real world settings than the previous approaches in developing intelligent pavement management systems.