Given the increasing demand for high-quality strip steel, there is a pressing need to enhance the capability of surface defect detection. However, the irregular sizes of surface defects and the complex backgrounds present substantial challenges for existing methods to achieve efficient and effective defect detection. In the light of this, we develop a novel model entitled MCF (Multi-scale Context Fusion for Strip Steel Surface Detection) to overcome these challenges. Specifically, we first utilize the GridMask data augmentation technique to simulate occlusion in defects and bolster the model’s robustness. Then, we integrate a Dynamic Convolution and Attention (DCA) layer into the backbone network, merging dynamic convolutions with self-attention mechanisms to adapt to diverse input image sizes and effectively exploit latent feature information. Finally, by introducing the Fusion layer within the neck network, we intensify the fusion of feature information from various levels and dynamically weigh feature maps of differing resolutions. The extensive experiments demonstrate that MCF outperforms the state-of-the-art baselines on two publicly available datasets.