To address the issues of scarce small defect samples and poor detection accuracy in magnetic flux leakage (MFL) testing of oil and gas pipelines, this paper proposes a small defect detection network for oil and gas pipelines on the basis of shallow feature suppression. First, an adversarial generative network is utilized, which incorporates prior knowledge to generate high?quality small defect samples. Subsequently, a defect feature suppression module is introduced during the feature extraction process, which suppresses the semantics of large defects in shallow pyramid features, thereby enhancing the features of small defects. Finally, a multi?scale Transformer is employed to fully leverage the local details and global information of the feature images, improving the accuracy of pipeline defect detection. The experimental results demonstrate that the accuracy of this model is 95.1%, which is 7.8% higher than the average value of existing faster R?CNN and other methods methods.