Improved YOLOv11 Algorithm for Highway Potholes Detection in Aerial Images by UAV

Addressing the challenges of significant multi-scale variations in target objects and insufficient detection accuracy caused by complex scenarios in UVA aerial image-based pothole detection on roads, this paper proposes an improved YOLOv11 algorithm for pothole detection in UVA-captured images. Firstly, in the backbone network, the original C3K2 features extraction module is replaced with a lightweight enhanced detection module (LEDM). By using a grouped parallel processing method, this module segments multi-scale pothole feature channels and dynamically extracts key pothole information through adaptive feature enhancement. It combines lightweight computation with the elimination of redundant parameters to improve both the accuracy of pothole feature extraction and the operational efficiency of the model. Secondly, in the neck of the network, an enhanced multi-scale attention fusion module (EMSA) is introdcued to replace the original feature fusion method based on upsampling concatenation and convolution. This module improves the efficiency of cross-scale information transmission in scenarios where small pothole features are diluted and large pothole features are confused with the background. It achieves this by combining dynamic attention calibration of pothole feature weights, grouped spatial refinement of pothole edges, and residual feature fusion. Experimental results show that the improved model achieves m_AP@50 and m_AP@0.50~0.95 scores of 86.6% and 58.3% respectively, representing improvements of 5.74% and 11.69% over the baseline YOLOv11n model. The recall rate reaches 82.7%, a 19.68% increase compared to the baseline. The experimental results demonstrate that the proposed optimization strategies effectively improve the model’s detection ability for multi-scale potholes with weak features and reduce the miss rate in pothole detection tasks for highway images captured by UVAs.