高级检索

基于结构理解的手绘草图表格识别

FreehandSketch Table Recognition based on Structure Understanding

  • 摘要: 基于笔交互的手绘草图表格比基于WIMP界面范式的传统电子表格更易于促进用户思维交流和创造性交互工作,同时可以避免用户学习成本高和交互界面烦琐等问题. 然而,由于缺乏专门用于手绘草图表格识别的开源数据集,以及草图本身具有的模糊性、抽象性和用户在绘制草图时的随意性,手绘草图表格的识别仍然面临着巨大的挑战. 为了解决存在复杂单元格、结构框线与内容重叠、笔迹重描补笔等具体挑战, 提出了一种基于结构理解的手绘草图表格识别算法, 通过一个支持向量机(support vector machines, SVM)对笔划信息进行划分, 再使用多峰值检测算法与真实交点找寻和单元格特征属性检测结合, 实现复杂草图表格结构的识别. 客观评估结果显示, 与OCR基线算法相比, 基于结构理解的手绘草图表格识别算法在面向草图表格结构识别任务中的树编辑距离相似度(Tree-Edit-Distance-based Similarity, TEDS) 指标提升了13%以上; 专家评估结果显示, 基于结构理解的手绘草图表格识别算法在表格结构识别和将内容匹配到对应单元格2个评价维度中的识别效果均优于OCR基线算法的识别效果.

     

    Abstract: pen-based sketching tables are more likely to facilitate users' communication and creative interaction, avoiding the problems of high learning cost and cumbersome interaction interface that exist in traditional spreadsheet systems, compared to traditional spreadsheets systems based on the WIMP interface paradigm. However, the recognition of hand-drawn sketch tables still faces great challenges due to the lack of open source datasets dedicated to the recognition of hand-drawn sketch tables, as well as the ambiguity and abstraction of the sketches themselves and the arbitrariness of the users in drawing them. In order to solve the specific challenges such as the existence of complex cells, the overlapping of structural frame lines and content, and the redrawing and filling of handwriting, propose an algorithm for freehand sketch table recognition based on table structure understanding. A support vector machines (SVM) is used to classify the stroke information, and then the multi-peak detection algorithm is combined with the real intersection finding and cell feature attribute detection constructed in this paper to achieve the recognition of complex sketch table structure. The objective evaluation results show that the Tree-Edit-Distance-based Similarity (TEDS) of this algorithm in the sketch-oriented table structure recognition task is improved by more than 13% compared with the OCR baseline algorithm; the expert evaluation results show that this algorithm outperforms the OCR baseline algorithm in the evaluation dimensions of table structure recognition and matching the content to the corresponding cells.

     

/

返回文章
返回