<p>In industrial settings, few-shot steel plate surface defect detection poses a challenging task. The goal is to identify new defect categories with only a limited number of samples available. However, existing few-shot object detection (FSOD) methods often struggle with complex and variable defects, particularly those characterized by large-scale geometric deformations, high intra-class similarity, and strong background interference. To address these issues, we propose a novel FSOD method called DSG-FSOD. First, we design a novel backbone network, DCN-ResNet101. By introducing deformable convolutions, this network dynamically adjusts the position and shape of convolution kernels. This enables more effective capture of large-deformation defects and complex geometric structures, enhancing the model’s flexibility and accuracy. Second, we integrate the Squeeze-and-Excitation (SE) module into the DCN-ResNet101 architecture. This integration enables the network to place greater emphasis on valuable features while suppressing irrelevant responses. This improvement leads to more reliable feature representation and boosts overall detection performance. Third, we embed a Global Context (GC) module into the DCN-ResNet101 network, enabling it to capture comprehensive contextual information. This enhancement improves the model’s category discrimination capability and effectively reduces false detection rates. To validate the effectiveness of DSG-FSOD, we conduct extensive experiments on two steel defect datasets, NEU-DET and X-SDD. The experimental results demonstrate that DSG-FSOD outperforms state-of-the-art methods under both evaluation protocols. Compared with baseline models, DSG-FSOD achieves substantial improvements in detection accuracy. Specifically, it achieves a 6.4% improvement at 2-shot and 5.5% improvement at 10-shot compared to baseline models, demonstrating its effectiveness and generalization capability in practical steel defect detection scenarios.</p>