A model-driven classification and recursive segmentation method for automatic panel extraction from biological and medical papers

We present a novel method to automatically extract panels from figures in biomedical articles. Our method consists of figure (or panel) classification and panel segmentation. Figure classification determines the existence of photograph in a figure. A Gaussian model is constructed for photographs and plots. Figures and panels are evaluated based on the model to determine their class. If it contains photographs, an iterative panel-splitting process follows. This process continues until no further straight lines are identified in the subfigures. Experiments were conducted with 182 figures from 25 articles published in different journals. Despite vast difference between figures, our method successfully extracted both plots and photographs and was able to identify zoom-in views that are superimposed on the original photographs.