A dual-image fusion instance segmentation model for pavement patch detection

Abstract

Accurate patch detection is essential for reliable pavement condition evaluation and life cycle assessment. However, this task remains challenging due to variations in patch morphology, visual similarity to the background, and the limited availability of comprehensive patch datasets. This paper presents a novel patch detection method that pioneers the use of instance segmentation techniques to obtain more detailed patch information and fuses dual-image data from the Laser Crack Measurement System (LCMS) to capture richer features for enhanced precision. Furthermore, the proposed method goes beyond conventional approaches that focus solely on basic detection by incorporating a patch counting method, enabling accurate patch quantification and area measurement across different road section lengths. Experimental results show that the proposed patch detection model (FuPatch) outperforms baseline models while maintaining comparable efficiency. Additionally, the patch counting method effectively quantifies both the number and area of patches. These findings demonstrate that the developed model not only effectively detects patches but also provides detailed spatial insights and performs accurate patch counting, making it highly applicable for real-world pavement condition assessments.