Bridging Data Distribution Gaps: Test-Time Adaptation for Enhancing Cross-Scenario Pavement Distress Detection

Abstract

Automatic pavement distress detection using deep learning has revolutionized maintenance efficiency, but deploying models in new, unseen scenarios presents significant challenges due to shifts in data distribution. Traditional transfer learning requires extensive labeled data from the new domain, which is both time-consuming and costly. This paper proposes a test-time adaptation (TTA) framework that addresses feature distribution biases across different scenes, including differences in background, perspective, and environmental conditions. It adapts models at inference time without requiring additional labeled data, making it a promising solution for cross-scenario applications. The framework dynamically adapts the model to these biases by generating domain-specific prior knowledge, applying perspective correction, and generating global attention maps to reduce focus on irrelevant elements. We evaluate the framework on a cross-scene dataset that includes pavement images from three countries and four perspectives. In unsupervised settings, the TTA framework improves detection accuracy by 20.6%, achieving 93.09% of the accuracy obtained through transfer learning with 10,000 labeled images. Compared to traditional transfer learning, our framework reduces the reliance on high-quality labeled data while achieving similar performance gains. Experimental results also demonstrate the framework’s adaptability across various deep learning detection models, offering a scalable solution for rapid deployment and cross-scenario application of pavement distress detection systems.

Keywords

• pavement distress
• deep learning
• cross-scene transfer
• test-time adaptation