融合多尺度特征的文档版面分析算法OA

To address the challenges of complex document layouts and inefficient parsing of multimodal elements,we propose a cross-layer,multi-scale document layout analysis algorithm based on the Transformer-based end-to-end object detection network with dual at-tention mechanisms.Building upon the standard DETR architecture,the proposed algorithm incorporates a positional attention module and a feature pyramid network-weighted coordinate attention module.The positional attention module enhances the model's spatial structuring capabilities,enabling effective recognition of strongly spatially dependent elements such as table grids and headings.The feature pyramid network improves multi-scale feature fusion,boosting detection accuracy for small objects and elements of varying sizes.The weighted coordinate attention module enables the model to focus more precisely on regional features associated with critical document elements.Experiments conducted on the PubLayNet dataset demonstrate that the proposed algorithm outperforms mainstream methods in detecting various document elements,including text,headings,tables,and illustrations.It achieves an all mAP@0.5 of 96.8%,representing a1.6 percentage points improvement over the original DETR.The proposed algorithm not only enhances the accuracy and robustness of document element detection but also lays a solid foundation for downstream tasks such as multimodal semantic fusion.