Abstract: Objective With the frequent occurrence of pipeline theft via illegal branch connections and issues such as cracks and corrosion due to long-term service, the demand for automated pipeline defect detection has become increasingly urgent. However, existing methods primarily rely on manual interpretation of defect magnetic signal images, resulting in low efficiency and high false-positive rates. Method To address the technical bottleneck of limited and singular defect representations in the field of pipeline defect detection, a novel intelligent defect detection method based on multi-representational image interactive learning is proposed, driven by practical engineering application requirements. This method achieves high-precision and intelligent automated detection by utilizing a random input strategy of three different data representations: line-based, grayscale, and pseudo-color images. This strategy effectively mitigates layer space confusion while fully leveraging multi-dimensional feature information. Additionally, a specially designed CutMix image augmentation framework is incorporated, significantly improving the model's ability to detect localized defects. Based on this, an intelligent system supporting real-time detection via video streams is developed, with a multi-representation iterative re-inspection mechanism that notably reduces the rate of missed detections. This method combines deep learning detection networks with the multi-dimensional feature characteristics of permanent magnetic disturbance signals, breaking the limitations of traditional single-representation input methods. Results In testing experiments conducted on actual pipeline data from twelve in-service engineering pipelines, including those from detectors of sizes ϕ 508 and ϕ 813, the proposed method achieved a recall rate of 90.64%, precision of 89.92%, and mAP@0.5 of 91.91% in tasks related to defect detection such as pipeline theft at branch connections. Conclusion Compared with existing detection algorithms, the proposed framework demonstrates promising performance with low computational cost. Additionally, it provides stable and reliable automated detection in pipeline video streams without the need for manual annotation. Proposed method, as a dedicated model for pipeline defect detection, has been integrated into an extensively used pipeline magnetic signal visualization software system, supporting online analysis, reducing labor costs, and providing robust technical support for combating oil theft and ensuring safe pipeline operation.