Intelligent detection of underground pavement defects using an improved YOLO11 framework based on ground penetrating radar

B-scan images obtained by ground penetrating radar (GPR) often show underground defects characterised by weak reflections, strong noise interference and significant scale variations. To address these challenges, this article proposes an improved detection model, YOLO11n-HISA, based on YOLO11n (You Only Look Once 11 nano), which enables highly accurate identification of two types of underground defects: voids and cracks. The model optimises the network architecture in three aspects (feature extraction, feature modelling and multi-scale fusion) adapted to the characteristics of GPR images. The C3k2-HetConv module is introduced in the feature extraction phase to improve local structural representation while reducing redundant calculations. The C2PSA-IERB module was designed during the feature modeling phase to improve the joint modeling capability of the global environment and local details. The ASF-SDI module is integrated into the neck structure to enhance consistency and complementarity between features at different scales. Experimental results obtained from a self-constructed set of underground pathological GPR data demonstrate that the proposed YOLO11n-HISA model improves overall precision, recall, and mean average precision (mAP) measurements while maintaining a low number of parameters and computational requirements. More specifically, mAP@0.5 increases from 0.809 to 0.857 and mAP@0.5:0.95 increases from 0.442 to 0.483, confirming the effectiveness of the algorithm.