FCDM‐Net: A Feature‐Calibrated Dynamic Masking Network Based on Pre‐Trained Models for Few‐Shot Catenary Component Anomaly Detection

ABSTRACT

Existing detection systems for electrified railway catenary components still face the following challenges. (1) The background of aerial images of the catenary support component is complex; (2) fault samples are difficult to obtain, and the significant feature shift between training and testing images results in a scarcity of negative samples available for training; and (3) the coexistence of microscopic texture defects and macroscopic logical anomalies in catenary components. To address these challenges, this paper proposes a novel feature‐calibrated dynamic masking network (FCDM‐Net) for the unified few‐shot anomaly detection of catenary components. First, a lightweight residual adapter is integrated with a frozen visual encoder. Through a few‐shot fine‐tuning strategy, the general visual features are mapped to the specific manifold space of catenary components, thereby achieving the feature domain calibration. Second, we introduce a noise‐resistant mechanism named heatmap‐guided dynamic masking, which employs a coarse anomaly heatmap extracted from the adapted features as prior guidance to dynamically filter background redundancy in the initial mask and accurately extract the component body. On this basis, a dual‐branch inference system is proposed to enable the collaborative detection of microscopic texture defects and macroscopic logical anomalies. Finally, experiments conducted on the real catenary component dataset collected by drones demonstrate that under a four‐sample setting, the proposed method achieves image‐level metrics (I‐AUROC (area under the receiver operating characteristic curve)/average precision (AP)/F1) of 94.7/93.1/96.7, respectively.