High-precision pothole detection using the ECC-YOLO network with deformable convolution and attention mechanisms

Road potholes present a significant challenge to urban traffic safety and infrastructure maintenance. Traditional manual inspection methods fail to meet the demands for real-time performance and high accuracy. In this study, we propose ECC-YOLO, a lightweight object detection model based on YOLOv11n, specifically designed for pothole detection in complex road environments. First, we introduce the C3k2DCY module, which extracts multi-scale features and leverages deformable convolution to enhance the model’s ability to capture irregular geometric structures of potholes. Next, a Contrast-Driven Feature Aggregation (CDFA) module is designed to improve feature discriminability at boundary regions by reinforcing the contrast between potholes and surrounding backgrounds, thereby significantly boosting detection precision. Furthermore, an Edge-aware Lightweight Attention-based Spatial Feature Pyramid Network (ELA-HSFPN) is integrated to enable effective fusion and semantic enhancement of multi-level features, jointly combining low-level edge details and high-level semantic cues for improved localization and classification of pothole targets. Experiments conducted on a custom pothole dataset demonstrate that ECC-YOLO achieves an accuracy of 84.5%, a recall of 67.9%, and a [email protected] of 74.2%, while maintaining real-time performance. Compared to the baseline YOLOv11n model, ECC-YOLO improves accuracy by 1.7% points, recall by 2.5% points, and [email protected] by 1.4% points. Ablation studies further confirm the individual contribution of each module to overall performance. Overall, ECC-YOLO demonstrates excellent capability in enhancing detection accuracy, reducing false positives, and adapting to complex environments, indicating strong potential for real-world deployment.