An improved official document layout element detection transformer based on RT-DETR
Bole Tan, Hong Xiao · 2025
To meet the practical needs of automated document review for governments and enterprises, we focus on layout elements detection in automated Chinese official document review tasks. By improving upon RT-DETR, we propose a document-level object detection model named Official Document Layout Element Detection TRansformer (ODLEDTR) that balances accuracy and efficiency. A specialized Official Document Layout Element Detection (ODLED) dataset compliant with the national standard Layout key for official document of Party and government organs is constructed, where challenges related to text sensitivity and stringent format constraints in governmental and entrepreneurial documents are resolved using the Virtual Content with Correct Format (VC-CF) generation strategy. Key improvements based on RT-DETR include: employing StarNet as the backbone to increase implicit dimension of features via elementwise multiplication and reduce parameters; using convolutional additive self-attention to linearize intra-scale feature interaction computational complexity; introducing Manhattan distance priors and efficient sample methods to enhance the global modeling ability of cross-scale feature fusion. Experimental results show that ODLEDTR reduces parameters by 49.3% and boosts inference speed by 50.3% compared to RT-DETR. It achieves the highest accuracy on both ODLED and similar domain dataset, surpassing mainstream models. ODLEDTR provides an achievable solution for automated document review systems.