NADC-YOLO: Neighborhood Attention Dynamic Kernels for Document Layout Analysis

Document layout analysis (DLA) is the foundational step in intelligent document analysis, as it identifies the structural regions enabling downstream tasks like OCR, information extraction, and semantic understanding to operate accurately on the correct content zones. However, applying DLA to real-world documents with diverse and complex structures remains a non-trivial task, particularly when both high detection accuracy and lightweight deployability are required. To address the challenge of balancing lightweight deployment and high-precision detection in complex document layout analysis, we propose NADC-YOLO, an enhanced YOLOv10-based model for document layout detection. Specifically, we first introduce Global Context-guided Neighborhood Attention Dynamic Convolution into YOLOv10 to enhance the representation capability of complex layout elements and structural relationships. Second, an asymmetric bidirectional feature fusion network is designed to improve the integration of multi-scale semantic information and localization details while maintaining a low parameter count. In addition, we construct ALCAI, a fine-grained Chinese document layout analysis dataset derived from government accident investigation reports. The dataset comprises 23 categories of layout elements, with an emphasis on fine-grained distinctions among paragraph types. Experimental results show that NADC-YOLO achieves competitive detection performance on datasets ALCAI, D4LA, and DocLayNet. Specifically, it achieves 93.0%[email protected] and 73.1%[email protected]:0.95 on the ALCAI dataset, while also delivering consistent improvements over the baseline model on datasets D4LA and DocLayNet. The proposed modifications yield substantial performance gains with only a modest increase in parameter count, demonstrating improved parameter efficiency, providing a reliable visual parsing foundation for industrial document understanding, information extraction, and retrieval-augmented generation applications.