To address the difficulty of balancing detection accuracy in complex backgrounds and edge-side computational efficiency in UAV-based inspection of surface corrosion defects on fly ash steel silos, an improved lightweight corrosion detection model YOLO-MCD based on YOLO11n was proposed. MobileNet V3 was used to reconstruct the backbone network and the SE (Squeeze-and-Excitation) channel attention mechanism was integrated to reduce computational redundancy during feature extraction and dynamically enhance the corrosion related feature response. By embedding a CAA (Context Anchor Attention) feature enhancement module at the junction between the backbone and the neck network, the network''s ability to perceive irregular corrosion boundaries and hidden textures was strengthened. A lightweight detection head DG_head combining Detail-Enhanced Convolution (DEConv) and GroupNorm (GN) with cross-level weight sharing was designed to reduce the parameters and computational overhead caused by independent multi-branch computation while maintaining the representation capability of corrosion detail features. The MPDIoU loss function is introduced to optimize the bounding box regression process, providing more accurate regression gradients to improve the localization capability for overlapping and complex shape targets. The experimental results show that the FLOPs of the model on the Corrosion inspection dataset Corrosion is reduced by 50.3%, mAP@0.5 reaches 88.86%, and increased by 3.62% compared with YOLO11n. The recall rate reaches 81.94%, and increased by 6.73%. The overall performance is better than the benchmark YOLO11n and other lightweight detection models, providing an efficient and lightweight detection solution for UAV-based inspection of surface corrosion defects on steel silos.