Robustness Evaluation of a JEPA-Based Model and YOLO for Pavement Distress Classification under Noise Corruptions

Abstract

While automated pavement distress classification is critical for Intelligent Transportation Systems (ITS), conventional evaluations on pristine datasets often mask deep learning vulnerabilities to real-world environmental degradation. This study evaluates architectural robustness by contrasting a supervised Convolutional Neural Network (YOLOv11m-cls) with a self-supervised Vision Transformer framework (StoP-JEPA). Models were subjected to extreme Out-of-Distribution (OOD) environmental stress tests, including Defocus Blur, Gaussian Sensor Noise, and Salt-and-Pepper impulse corruption. Furthermore, a Balanced Performance Index (BPI) is introduced to quantify diagnostic symmetry between fundamentally distinct crack topologies. Despite near-perfect baseline accuracy, stress testing revealed distinct architectural inductive biases. YOLO demonstrated inherent robustness to optical blur via local convolutional low-pass filtering, whereas the StoP-JEPA configuration utilizing Global Average Pooling (GAP) excelled under Gaussian noise through statistical smoothing. Crucially, under extreme impulse noise, GAP suffered catastrophic asymmetric collapse, dropping to a BPI of 7.06% (± 4.67%). Replacing GAP with an attention-driven [CLS] token effectively mitigated this vulnerability. By selectively routing semantic context from uncorrupted patches, the [CLS] mechanism maintained a statistically stable BPI of 61.71% (± 3.94%). Ultimately, for practical ITS deployment, StoP-JEPA provides the most stable diagnostic equilibrium under severe Gaussian noise, while YOLO remains superior under optical blur, highlighting the necessity of aligning architectural priors with specific deployment environments.