A Masked Multi-View Bidirectional Network for Class Extremely Imbalanced Object Detection Under Magnetic Flux Leakage Signals

Magnetic flux leakage (MFL) detector can analyze the state of pipelines and is widely used in pipeline inspection. Object detection of class extremely imbalanced MFL signals is a challenging and demanding task. It is difficult for existing deep learning-based detection methods to effectively detect a small number of component classes. To address the above issues, a novel masked multi-view bidirectional network (MMB-Net) is proposed in this paper. First, based on the problem of insufficient component samples, a mask-based data augmentation module is designed, where we develop a novel wavelet convolution blocking module and Gauss-like distribution mask strategy so as to expand component sample sets. Second, based on the problem that poor feature extraction of partial MFL signals leads to low accuracy of component detection, a learnable multi-view attention (LMA) module is designed to expand the perceptual field of the network and fully mine the features of MFL signals. Finally, a bidirectional attention module based on multiple classification layers (BAMC) is proposed to learn different decision boundaries, which maintains the accuracy of the defects while improving the accuracy of the components. Experimental results illustrate that the proposed network can effectively build a detection model under MFL signals and outperforms the state-of-the-art methods on mean precision (9.5%).
Note to Practitioners
—This article is motivated by the problem of extreme class imbalance in the MFL signals collected from pipelines, which seriously reduces the detection accuracy of defects and components in the MFL signals. In terms of the MFL signal intrinsic characteristics, a masked multi-view bidirectional network (MMB-Net) based on data dynamic adjustment of model parameters and bidirectional learning decision boundaries are constructed. Our method can extend the component samples and learn different decision boundaries to improve the detection accuracy of the component classes. The experimental results show that our proposed method is more advantageous than the state-of-the-art methods, and therefore, our proposed method has strong practical value.