A Novel Geometry-Aware Context Selection Network for Small Defect Detection Under Magnetic Flux Leakage Signals

Defect detection is crucial for evaluating the pipeline safety condition. However, due to the limited size of small defects and strong background interference in complex scenarios, the performance of current defect detection methods sharply declines in detecting small defects. To address the above issues, this paper focuses on detecting small defects via fine-grained geometry-aware feature extraction and selective feature fusion strategy, which explicitly enhances the defect feature from multiscale contexts. Specifically, a novel shallow geometry-aware feature extraction module (SGFEM) is integrated with a context selection module (CSM), exploiting the complementary information of shallow texture features and deep semantic representations for small defect detection. Moreover, the double supervision loss compensates for feature degradation and information loss while optimizing the decision process of the prediction model through the multi-level prior supervision (MPS) module, so that the model pays attention to small defects. Our method has been validated through real pipeline and industrial application cases. The results demonstrate that the detection accuracy of our proposed method outperforms the state-of-the-art methods on the pipeline defect dataset (AR and AP increased by 5.9% and 2.4%), and the proposed method has a high industrial application value.