Surface defect detection is a critical technology in the field of intelligent manufacturing, playing a significant role in improving product quality. The core task of typical object detectors is to achieve defect classification and localization through bounding box annotations. However, due to the multi-scale nature of surface defects and the complexity of the background, the inherent information sparsity in the supervision signals of bounding box labels often makes it difficult for models to accurately locate defect positions and identify the categories. To address this issue, we propose a Bounding Box-Derived Mask Guidance Network for Accurate Surface Defect Detection (BDMGNet). During the training phase, BDMGNet first generates Bounding Box-Derived Masks (BDM) from object bounding boxes and enhances the backbone network’s localization capability for object regions by incorporating an Auxiliary Segmentation Guidance Subnetwork (ASGS). Additionally, a Preference Dynamic Weight Averaging (PDWA) strategy is introduced to optimize the model’s attention allocation for detection tasks. During the inference phase, the ASGS can be pruned and thus improve defect detection accuracy without increasing the original model’s parameter count or complexity. Notably, the ASGS exhibits high generality and adaptability, functioning as a plug-and-play module for all object detection networks while maintaining the original model’s inference speed and effectively improving accuracy. Experimental results on the NEU-DET, DeepPCB, and GBC-DET datasets demonstrate that BDMGNet achieves significant performance improvements across series baseline models, validating the effectiveness and universality of the proposed method.