Improved Leaf Image Classification Using Topological Features and CNN With Attention Module

Abstract

Leaf image recognition is a fine-grained image classification task in the computer vision domain that is fundamental yet challenging both in biology and computer vision research communities. With the advent of the Convolutional Neural Network (CNN), encouraging accuracy has been achieved on leaf image recognition. However, the current CNN-based methods mainly capture the leaf properties of geometric and statistical distribution aspects but ignore the topological features. In this paper, we integrated the topological descriptor Persistence Image (PI) with the classic CNN model VGG16 to improve the accuracy of leaf image classification. A pretrained model Pi-net was adopted as the PI extractors, and an attention module was introduced to fuse the topological features and the features learned by CNN. Experiments on species dataset Flavia, Swedish, and Folio, and a more fine-grained cultivar dataset Cherry demonstrated the effectiveness of the proposed method. Our method achieved an accuracy of 99.85% on the Flavia dataset, 99.92% on the Swedish dataset, 99.64% on the Folio dataset, and 68.95% on the Cherry image dataset. Besides, the huge accuracy improvement on the Cherry dataset also demonstrated the advantage of our method in fine-grained leaf image classification.