Texture image classification based on persistent homology

In the field of computer vision and image processing, texture provides critical visual clues about the composition of internal regions of an image. This paper proposes a novel texture image classification method using persistent homology (PH) theory from topological data analysis (TDA) to extract topological information of texture images at different scales and dimensions. We convert texture images into grayscale images and use their pixel intensity values as filtration thresholds, tracking changes in 0-dimensional and 1-dimensional persistent homology classes under different thresholds to generate persistence diagrams of the texture images. The persistence images are then converted into feature vectors. Our method is validated with 10-fold cross-validation using a random forest model on the Kylberg dataset and KTH-TIPS dataset, achieving accuracies of 99.84% and 87.90%, respectively.