HPC-enabled Deep Learning Multi-model for Road Damage Detection

Abstract. Predictive maintenance for road infrastructure is a way-forward approach that replaces traditional reactive and routine maintenance with insights derived from AI. Accurate prediction of surface defects facilitates timely intervention, prioritizes repair effort, and optimizes cost. However, it is challenging due to lack of appropriate labelled dataset, larger data volume, photographic angle, poor resolution and surface types. In this paper, we propose a deep-learning pipeline that seamlessly integrates task-specific models for detection of pothole and waterlogging along with a crack detection model trained on a subset of publicly available RDD2022 dataset. GhostConv was adopted to enhance the YOLOv11 for crack detection. The model was trained on 640 x 640 resolution image with standard augmentations like flipping, scaling, hue and saturation using C-DAC PARAM Siddhi-AI HPC system. The augmentation techniques improved the model robustness across varied light and surface conditions. This multi-model approach allows new damage categories to be incorporated easily by retaining individual components without changing the core system. The results indicate that the multi-model pipeline is capable of detecting road damages such as pothole, waterlogging, longitudinal, transverse and alligator cracks from images. Our output model can be seamlessly integrated with online geospatial applications such as GeoSevak and GeoSadak (PMGSY National GIS), that can result in incredible increase in the implementation efficiency and transparency.