Topological data analysis with digital microscope leather images for animal species classification

This study presents a method for classifying cow and horse leather using a small number of digital microscope images and topological data analysis. In this method, hair pore coordinates in the images are used as essential information for classification. First, the coordinates were semiautomatically extracted using conventional image processing methods and persistent homology (PH) computation. Binary images with white pixels corresponding to the coordinates were generated, and their PHs were computed using filtration based on the Manhattan distance. In addition to the pairwise distance between the two pores, zeroth- and first-order lifetimes were used as explanatory variables to construct the classifier. Among the three explanatory variables, the zeroth-order lifetime resulted in the highest classification accuracy (86%) for the test data. Furthermore, we constructed logistic regression (LR) and random forest (RF) models using the zeroth-order lifetime computed from all images and conducted model interpretation. In both LR and RF, information on a zeroth-order lifetime of less than 10 was used as an important explanatory variable. Additionally, the inverse analysis of birth-death pairs suggested that the zeroth-order lifetime contains topological information distinct from the conventional pairwise distance. Our proposed method is designed to be robust in data-limited situations because it only uses hair pore coordinates as explanatory variables and does not require other information, such as hair pore density or pore size. This study demonstrates that accurate classifiers can be obtained using topological features related to hair pore arrangement.