为提高管道缺陷图像检测的准确率，提出了一种基于改进RefineDet的管道数字射线成像（digital radiography, DR）缺陷图像检测模型。该模型针对管道DR缺陷图像数据少、目标少等特点，从以下三个方面进行改进。首先，在骨干网络设计方面，使用Swin transformer代替VGG16作为主干网络，在提高特征提取能力的同时减少主干网络参数量。其次，针对管道DR缺陷图像目标数量较少而易受背景干扰问题，通过在主干网络与特征融合阶段之间加入全局注意力模块来强化模型对重要特征的关注，从而提高检测性能。最后，在后处理阶段，针对传统的非最大值抑制算法直接去除非最好预测框问题，使用软非最大值抑制算法以更合理的方式去除非最优预测框。结果表明：该方法能够有效实现管道DR缺陷图像的检测，并且相比于其他4种常用的目标检测模型，提出的模型可以有效提升管道DR缺陷图像检测的准确率，研究结果可为DR缺陷图像检测提供技术支撑。

To improve the accuracy of pipeline defect image detection, a pipeline digital radiography defect image detection model based on the improved RefineDet was proposed. This characteristics of limited pipeline DR defect image data and the scarcity of targets were addressed by making improvements in three aspects. Firstly, in the design of the backbone network, Swin Transformer was used instead of VGG16 as the backbone network, which enhanced the feature extraction capability while reducing the number of parameters in the backbone network. Secondly, to address the problem of limited targets in pipeline DR defect images and vulnerability to background interference, a global attention module was introduced between the backbone network and the feature fusion stage to enhance the model's focus on important features, thereby improving detection performance. Lastly, in the post-processing stage, a soft non-maximum suppression algorithm was used to remove non-optimal predicted boxes in a more reasonable way, as opposed to directly discarding non-maximum predicted boxes using traditional non-maximum suppression algorithms. The results show that the proposed method can effectively detect pipeline DR defect images. By comparing with four other commonly used object detection models, the proposed model significantly improves the accuracy of pipeline DR defect image detection. The research results can provide technical support for the detection of DR defect images.