基于YOLOv10的事故调查报告版面分析方法研究

Research on Accident Investigation Report Layout Analysis Method Based on YOLOv10

摘要: 近年来，随着自然和人为灾害的频发，事故调查报告日益彰显其重要性。针对事故调查报告版面布局复杂、元素种类繁多且尺度差异大，导致现有模型在关键信息提取中出现检测框重叠和小尺度元素识别困难等问题，本文提出了一种基于改进YOLOv10的事故调查报告版面分析方法。首先，通过收集整理各省应急管理部门发布的文档，构建了包含2500张图像的事故调查报告专题数据集，并定义了23种细粒度布局元素标签以增强语义区分度。其次，在YOLOv10主干网络中引入GL-CRM模块，通过动态分配计算资源，增强模型对不同尺度目标的特征提取能力。同时，利用YOLOv10的无非极大值抑制(NMS-free)策略，减少检测框重叠并提高推理速度。实验结果表明，改进后的模型在专题数据集上的F1分数和mAP分别达到了87.72%和88.5%，相较于基线模型分别提升了约4%和7%。验证了该方法在事故调查报告文档智能化处理中的有效性和优越性。

Abstract: In recent years, with the frequent occurrence of natural and man-made disasters, the importance of accident investigation reports has become increasingly apparent. To address the challenges of complex layouts, diverse element types, and significant scale variations in accident investigation reports—which lead to overlapping detection boxes and difficulties in recognizing small-scale elements with existing models during key information extraction—we propose a layout analysis method based on an improved YOLOv10. First, by collecting documents published by provincial emergency management departments, we construct a specialized dataset for accident investigation reports containing 2500 images, and 23 fine-grained layout element labels were defined to enhance semantic discrimination. Second, a GL-CRM module is introduced into the YOLOv10 backbone network. By dynamically allocating computational resources, this module enhances the model’s feature extraction capabilities for targets of different scales, thereby resolving the issue of feature loss in fine-grained elements. Additionally, utilizing YOLOv10’s NMS-free strategy, a dual-head architecture (comprising one-to-many and one-to-one detection heads) is adopted to reduce detection box overlap and improve inference speed. Experimental results indicate that the improved model achieved an F1 score of 87.72% and an mAP of 88.5% on the self-constructed dataset, representing improvements of approximately 0.04 and 0.7, respectively, compared to the baseline model. These results validate the effectiveness and superiority of the proposed method in the intelligent processing of accident investigation report documents.