A Multi-Scale Hybrid Segmentation and MRMR Feature Selection Approach for Road Crack Detection Using XGBoost Classifier

Road crack detection plays an important role in the latter, which is crucial to road safety and infrastructure maintenance.This paper proposes a novel architecture embedding deep features extracted by an Improved PSP-Net and Enhanced HED Network (HED), together with Local Binary Pattern (LBP)-based texture features.The Improved PSP-Net collects multiscale contextual information in different scenarios for correct object detection, and the Enhanced HED Network enhances edge detection to more accurately classify crack boundaries.Such use of LBPbased texture features enriches the model's ability to capture small details of surface unevenness which helps in detecting intricate and complex crack morphology.In addition, it uses a Minimum Redundancy Maximum Relevance (MRMR) feature selection technique that selects the relevant and non-redundant features which increases the efficiency.A sophisticated XGBoost classifier then assigns the detected cracks into specific classes, allowing for robust and accurate crack type prediction.Experimental evaluations demonstrate that the model outperforms (99.09%) and is more robust (F1-score 99.11%, sensitivity 99.53%) than both conventional image processing methods (Gabor filtering + morphological methods) and deep learning methods (CNNs, ConvNets, optimized YOLOv8, Inception V2, DeepCrack, CrackSeg, U-HDN), suggesting that the proposed model holds wide potential impact on actual road maintenance and inspection applications.