Zero-shot framework for automated inspection of bridge expansion joints using road scan images

Monitoring bridge expansion joints using road scan images remains a challenge due to the substantial requirement for labeled data and model training inherent in traditional supervised learning approaches. The zero-shot learning framework that does not require additional model training by users has potential for realizing an effected monitoring system for bridge expansion joints. The proposed framework integrates Grounding DINO and the Segment Anything Model (SAM) with an automated point-prompt selection algorithm and a mode-based estimation technique. Experimental results demonstrate that the framework accurately detects and segments joint gaps, achieving an intersection over union of 0.92 and a Dice coefficient of 0.96, comparable to or surpassing those of supervised models. These findings indicate that the zero-shot approach enables reliable and precise measurement of joint gaps while eliminating the burden of data labeling and retraining. Future research will focus on developing fully automated prompt generation mechanisms and validating the framework under diverse imaging conditions to improve robustness and generalizability.