Magnetic flux leakage (MFL) internal detection is the core technology of pipeline internal detection, which is crucial to ensuring the safe transportation of pipelines. Due to the long-term underground or deep sea environment of pipelines, there are many small defects on the surface of pipelines. Due to the limited information available on small defects, traditional deep learning defect detection methods have difficulty achieving satisfactory detection results for small defects. A composite backbone network-based signal detection method for small magnetic leakage defects is proposed. First, a data enhancement method called background compression is proposed to compress background signals and thus enhance key features of small defects. Secondly, an adaptive positive and negative sample allocation strategy is designed to address the issue of uneven positive and negative sample allocation for small defects in the region proposal network. Finally, a multi-branch high-resolution feature extraction network for small defects is proposed, which uses a multi-branch composite structure to obtain high-resolution features for feature fusion, thereby improving the network's utilization of small defect texture information. The proposed method is validated using pipeline data from a test site, and the experimental results show that the proposed method is effective, achieving a detection accuracy of 90.3%, with an 8.4% mAP improvement compared to the best results.