To address the critical challenges in steel surface defect detection, such as high model complexity, frequent missed detection of small targets, and interference from complex backgrounds, we propose DCG-DETR, a novel detection framework based on the real-time detection transformer (RT-DETR) architecture, which leverages dual-branch feature enhancement and context-guided optimization. First, we design a dual-branch feature enhancement module that integrates an efficient channel attention mechanism with dynamic receptive field convolution in parallel, effectively enhancing the discriminative power for micro-defects against complex textures. Second, we construct a content-context guided aggregation feature pyramid (CCGAFP) module, which adopts content-aware upsampling CARAFE and global-local dual-path fusion to resolve multi-scale feature misalignment and improve the localization accuracy of small defects. Furthermore, we introduce a lightweight feature fusion module VoV-GSCSPC (generalized cross stage partial network with GSConv) that compresses redundant computations and optimizes cross-stage gradient propagation, significantly reducing model complexity while maintaining detection performance. Experiments on the NEU-DET dataset demonstrate that the DCG-DETR achieves 81.5% mAP@0.5, outperforming the baseline RT-DETR-L by 3.2%, while reducing parameters by 11% and computational load by 25.9%. The results validate that the DCG-DETR achieves a superior balance between accuracy and efficiency, offering a practical and deployable solution for industrial quality inspection.