High-accurate defect and component detection are essential to ensure the safe operation of pipelines. Aiming at the difficulties of low accuracy and poor generalization of existing detection methods, this paper proposes a novel pipeline defect and component detection method based on a heterogeneous diffusion model. First, the raw magnetic flux leakage (MFL) signal is pre-processed to reduce the influence of negative factors such as noise in the signal acquisition. Second, to address the difficulty of feature extraction, a feature extraction method based on sparse attention module is designed, which establishes the long-distance dependence relationship between the MFL signals through sparsification and then enables the model to focus on the information of defects and components. In addition, the traditional feature pyramid network is replaced by the path aggregation feature pyramid network, which fully ensures the completeness of multi-scale features. Finally, a detection mechanism based on the heterogeneous diffusion model is designed, which converts the candidate frame regression process into a denoising process for random frames, which reduces the model's dependence on a predefined anchor window, and in turn improves the model's generalizability and accuracy. The experimental results show that the average detection accuracy of the proposed method reaches 97.4%, which is better than the state-of-the-art comparative method by 3.5%, which ensures the prospect of its practical application.