Multi-task Lightweight Network Learning for Substation Equipment Recognition and Localization

With the growing demand for the recognition and localization of substation equipment, deep learning-based object detection algorithms have made significant progress. However, high model complexity and single-task detection remain major obstacles to the widespread application of these algorithms. To address these issues, this paper proposes a multi-task lightweight network learning algorithm for substation equipment recognition and localization, which combines object detection with tasks such as bounding box regression and keypoint detection in a joint training process. Based on RT-DETR, MobileNetV4 is selected as the lightweight backbone network, reducing model parameters while fully extracting image features. Additionally, a lightweight and efficient hybrid encoder is designed, consisting of an intra-scale feature interaction module that focuses on contextual information to effectively extract keypoint features, and a lightweight cross-scale feature fusion module. Finally, a Transformer-based prediction head is used for decoding. The performance of the proposed method is significantly improved compared with other comparison methods.