MTA-Net: A One-Stage Detector Based on a Multiscale Task-Aligned Network for Catenary Support Components

In the service status monitoring system of catenary support components (CSCs), the localization and detection performance for CSCs are directly related to the performance of downstream tasks such as anomaly identification and defect recognition. However, the CSCs detection faces problems such as multiple categories, multiscales, and small scales, which makes it difficult for general object detection algorithms based on deep learning to effectively exert their detection performance. To address the above problems, this article proposes a new one-step detector based on a multiscale task-alignment network for detecting the CSCs. First, the designed depth-wise separable residual atrous spatial pyramid pooling (DRASPP) module and deformable convolution (DCN) are introduced into the RegNetX-400MF backbone to optimize its capability of adaptability for multiscale objects; Second, the neural architecture search (NAS) and cross-layer feature balancing modules are introduced into the feature pyramid network (FPN) to improve the effect of cross-layer feature fusion and the detection effect for small-scale objects; Finally, a deformable task-aligned detection head (DTA-Head) and the corresponding task-aligned learning strategy (TAL) are introduced in the proposed method to further optimize the comprehensive detection performance of the model. Experiments show that the modules and learning strategies proposed in this article are effective, reasonable, and have certain advantages. Moreover, the proposed detection framework can achieve a detection accuracy mean average precision (mAP) of 49.53% on CSCs dataset while maintaining low computational complexity. Therefore, the method proposed in this article can be effectively applied to CSCs detection.