Traditional image processing techniques face many challenges in corrosion detection, such as complex backgrounds and obstruction interference, and are highly sensitive to lighting conditions, lighting angles, and environmental changes. Deep learning has good adaptability for object detection in complex environments and shows significant advantages compared to other methods. To effectively identify the corrosion situation on the surface of coastal public facilities, the CBG-YOLOv5s model was proposed for facility surface corrosion detection. This model can automatically identify and locate corrosion areas on metal surfaces and classify them into three corrosion levels based on the degree of corrosion: mild, moderate, and severe. Firstly, the YOLOv5s model C3 module was integrated with the convolutional attention mechanism CBAM, and the C3CBAM module was designed to enhance the channel and spatial attention capabilities of the feature map, thereby improving the feature expression effect. Secondly, based on the idea of multi-scale feature fusion, a small target detection layer is added to the feature fusion section to improve the detection performance of small targets. Then, a lightweight C3Ghost module was designed to reduce the model's parameter and computational complexity and improve the model's running speed. Finally, a dataset containing 6000 typical images of metal surface corrosion was constructed to validate the effectiveness and rationality of the proposed model. The experimental results show that the detection accuracy and speed of the CBG-YOLOv5s model are superior to YOLOv5s and other mainstream object detection models.