<p>Bridge surface crack detection is essential for ensuring bridge transportation safety and maintenance planning. Under complex backgrounds, models are easily interfered by stains, fine cracks, and concrete texture, leading to low detection accuracy and missed detections. Therefore, we propose an improved bridge crack detection algorithm, EL-YOLOv8. Firstly, the original YOLOv8n model has insufficient crack feature extraction capability under complex background interference. To address this, we introduce an efficient channel attention (ECA) fused with the C2f module to enhance the model’s ability to extract crack edge features in strongly disturbed backgrounds. Secondly, during multi-scale feature extraction, the original model inadequately preserves fine crack information in the feature maps, making it susceptible to noise and irrelevant details. Consequently, we innovatively designed a new feature map processing module that uses SCConv to filter redundant information in the feature maps, strengthening the feature representation of crack regions. Combined with the CSP structure, this better retains high-resolution features of fine cracks, addressing the original model’s weakness of severely missing fine crack detections. Finally, in bridge crack detection tasks with numerous difficult samples and particularly complex backgrounds, the original CIoU loss function pays insufficient attention to crack regions in hard samples, resulting in poorer bounding box localization accuracy. To solve this, we introduce the WIoU loss function, which assigns higher gradient gains to low-quality predicted boxes, reinforcing the model to learn more robust features. The experimental results demonstrate that compared to the baseline model, the EL-YOLOv8 model achieves improvements of 2.8% in precision, 2.7% in recall, and 2.3% in mAP@0.5. With an FPS of 15 on the Jetson Nano B01, it meets the real-time requirements for edge computing applications such as bridge inspection robots.</p>