Synergistic Enhancement: A Study on the Design of Large Models Assisted by End-to-End Road Damage Prompt Network and Methods for Quantification of Damage Morphological Features

Pixelwise segmentation and accurate quantification of road damage are crucial for road maintenance. However, the diversity of road damage types and forms, coupled with the high cost of semantic segmentation annotation, presents challenges for achieving precise pixelwise segmentation and morphological quantification. A novel end-to-end prompt-assisted network for road damage segmentation is proposed, integrated with a large model designed for quantifying morphological features. This integration enables pixel-level segmentation and quantification under zero-shot conditions. The proposed model is structured into preprocessing and postprocessing stages. An end-to-end damage box bounding prompt network (DBBPN) is introduced in the preprocessing stage, which leverages a combination of dynamic snake convolution (DSConv) and deformable attention (DAttention) mechanisms to focus on key areas of irregularly shaped damage. A Dy-detect head with scale, spatial, and task awareness modules is incorporated into this network, significantly enhancing its capability to detect complex types of road damage. A quantification method based on normal vector estimation and branch-point detection is proposed in the postprocessing stage, which substantially improves the accuracy of key parameters for quantifying various damage features. Evaluation on the RDD2022 dataset and a self-made validation dataset, MaskRDD-140, demonstrates that this method outperforms existing models in ${F}1$ -score, mAP50, and mean intersection over union (MIoU) metrics. Robust technical support is thus provided for efficient road maintenance.