LDA-YOLO11: a lightweight detail-attention detector for edge-deployable road surface defect detection

Abstract

Abstract Road surface defects, such as cracks, pavement damage and potholes, shorten road service life and affect driving safety and preventive maintenance. Existing lightweight object detectors are computationally efficient but still suffer from missed detections and unstable localization when dealing with elongated cracks, weak-texture damage and complex pavement backgrounds. To address these issues, this study proposes LDA-YOLO11, a lightweight detail-attention road defect detector based on YOLO11n. The model improves the neck feature fusion stage while keeping the original YOLO11n backbone and P3, P4 and P5 detection heads unchanged. A P2-guided Detail-Neck is designed to inject shallow high-resolution detail information into the P3 feature without adding an independent P2 detection head. Efficient Channel Attention modules are further introduced after the final P3 and P4 fusion features to recalibrate defect-related channel responses and suppress background interference. Experimental results show that LDA-YOLO11 achieves mAP@0.5 and mAP@0.5:0.95 values of 0.951 and 0.791, improving YOLO11n by 2.3 and 1.7 percentage points, respectively. On the Jetson Orin NX platform, TensorRT FP16 acceleration reduces the inference latency to 27 ms/image, corresponding to 37.03 FPS, while causing only a 0.2 percentage-point decrease in mAP@0.5. These results indicate that LDA-YOLO11 provides an effective and edge-deployable solution for real-time road surface defect detection in practical inspection scenarios.