Abstract
Crack sealant is commonly used to fill pavement cracks and improve the Pavement Condition Index (PCI). However, during asphalt pavement hot in-place recycling (HIR), irregular shapes and random distribution of crack sealants can cause issues like agglomeration and ignition. To address these problems, it is necessary to mill large areas containing crack sealant or pre-mark locations for removal after heating. Currently, detecting and recording crack sealant locations, types, and distributions is conducted manually, which significantly reduces efficiency. While deep learning-based object detection has been widely applied to distress detection, crack sealants present unique challenges. They often appear as wide black patches that overlap with cracks and potholes, and complex background noise further complicates detection. Additionally, no dataset specifically for crack sealant detection currently exists. To overcome these challenges, this paper presents a specialized dataset created from 1983 pavement images. A deep learning detection algorithm named YOLO-CS (You Only Look Once Crack Sealant) is proposed. This algorithm integrates the RepViT (Representation Learning with Visual Tokens) network to reduce computational complexity while capturing the global context of images. Furthermore, the DRBNCSPELAN (Dilated Reparam Block with Cross-Stage Partial and Efficient Layer Aggregation Networks) module is introduced to ensure efficient information flow, and a lightweight shared convolution (LSC) detection head is developed. The results demonstrate that YOLO-CS outperforms other algorithms, achieving a precision of 88.4%, a recall of 84.2%, and an mAP (mean average precision) of 92.1%. Moreover, YOLO-CS significantly reduces parameters and memory consumption. Integrating Artificial Intelligence-based algorithms into HIR significantly enhances construction efficiency.