DBF-RTDETR：基于DETR的工业产品表面缺陷检测

DBF-RTDETR: Industrial Product Surface Defect Detection Based on DETR

摘要: 工业产品表面缺陷检测作为智能制造的关键环节，亟需解决复杂纹理干扰、多尺度缺陷共存及实时性要求高等问题。基于此，本文提出一种基于双卷积–双向特征金字塔结构RTDETR (DBF-RTDETR)模型的表面缺陷检测的方法，旨在针对工业表面缺陷检测场景进行性能优化。首先，为增强模型对微小缺陷的特征捕获能力，设计了双卷积模块，其采用双分支并行结构，在降低计算量的同时增强了微小缺陷的特征表达能力；其次，构建了双向特征金字塔结构网络，通过双向跨层连接与多尺度特征融合，有效抑制了光照不均与背景纹理干扰；此外，通过引入多样分支块模块，有效提升了推理速度。最后，在工业产品数据集上进行了实验，验证了本文方法具有更好的表面缺陷检测效果，在提取精度、完整性、准确性等方面均有显著提升。

Abstract: Surface defect detection of industrial products, as a key link in intelligent manufacturing, urgently needs to address challenges such as complex texture interference, coexistence of multi-scale defects, and high real-time requirements. To tackle these issues, this paper proposes a surface defect detection method based on a Dual Convolution-Bidirectional Feature Pyramid RT-DETR model (DBF-RTDETR), aiming to optimize performance for industrial surface defect detection scenarios. First, to enhance the model’s feature capture capability for micro-defects, a dual convolution module is designed. This module adopts a dual-branch parallel structure, which reduces computational complexity while strengthening the feature representation ability for tiny defects. Second, a bidirectional feature pyramid network is constructed, which suppresses uneven lighting and background texture interference through bidirectional cross-layer connections and multi-scale feature fusion. Additionally, by introducing the Diverse Branch Block (DBB) module, the inference speed is significantly improved. Finally, experiments on industrial product datasets demonstrate that the proposed method achieves superior surface defect detection performance, with significant improvements in detection precision, integrity, and accuracy.