基于增强FPN的Vision Transformer在文档布局分析任务中的应用研究OA

Compared with traditional methods based on Convolutional Neural Networks(CNN),the document layout analysis model based on Vision Transformer can provide robust semantic and visual representations for downstream tasks through multi-modal pre-training mechanisms.However,the current multi-scale feature generation module and cross-resolution feature fusion process are prone to causing the loss of category attributes and boundary details,which in turn leads to issues such as category confusion and blurred boundaries.To address this bottleneck,Local Feature Enhancement Generation(LFEG)and Global-to-Local Feature Enhancement Fusion(GLEF)techniques are proposed to construct an enhanced Feature Pyramid Network(FPN)structure for achieving novel feature optimization.Specifically,the LFEG module optimizes four resolution modification modules to generate the multi-scale feature,while the GLEF module optimizes the traditional top-down fusion approach.Experimental results demonstrate that the proposed enhanced FPN structure can effectively improve the category consistency and boundary clarity of multi-scale feature maps,providing key technical support for optimizing the accuracy of document layout analysis based on Vision Transformer.