Research on Transformer Nameplate Rivet Recognition Based on Improved RT-DETR

To solve the difficulties of heavy effort and low efficiency in traditional manual collecting of transformer nameplate information, a transformer nameplate rivet target detection algorithm based on improved RT-DETR is proposed. This algorithm utilizes Partial Convolution (PConv) to replace some conventional convolutions, enabling feature extraction from only selected channels of the feature map, thereby reducing computational redundancy. Additionally, structural re-parameterization techniques are introduced to perform multi-branch training and single-path inference on the feature map, further decreasing computational costs. To solve the problem of tiny objects' poor detection accuracy, a novel feature fusion network architecture, ASF-DETR, is introduced. According to experimental data, the enhanced approach reduces the number of Calculation and parameter quantity by 16.9% and 27.5%, respectively, while increasing detection accuracy by 1.4% when compared to the original DT-DETR model. This improvement ensures real-time performance while effectively enhancing detection accuracy and meeting the lightweight requirements for field deployment.