A novel figure panel classification and extraction method for document image understanding

摘要

With the availability of full-text documents in many online databases, the paradigm of biomedical literature mining and document understanding has shifted to analysis of both text and figures to derive implicit messages that are unforeseen with text mining only. To enable automatic, massive processing, a key step is to extract and parse figures embedded in papers. In this paper, we present a novel model-driven, hierarchical method to classify and extract panels from figures in scientific papers. Our method consists of two integrated components: figure (or panel) classification and panel segmentation. Figure classification evaluates each panel and decides the existence of photographs and drawings. Mixtures of photographs and non-photographs are divided into subfigures. The splitting process repeats until no further panel collage can be identified. Detection of highlighted views is addressed with Hough space analysis. Using reconstruction from Hough peaks, enclosed panels are retrieved and saved into separate files. Experiments were conducted with a total of 360 figures extracted from two sets of papers that are retrieved with difference sets of keywords. Experimental results demonstrated that our method successfully segmented figures and extracted photographs and non-photographs with high accuracy and robustness. In addition, our method was able to identify zoom-in views that are superimposed on the original photographs. The efficiency of our method allows online implementation.