Hierarchical Classification of Animal Images Including Visually Similar Species: A Case Study on Parrot Images

In recent years, convolutional neural network (CNN)-based methods have gained significant popularity for image classification tasks. However, when faced with many classes, especially those with comparable features, CNN-based algorithms encounter difficulties in achieving effective differentiation. This limitation becomes particularly challenging when attempting to accurately classify animals and plants belonging to internationally endangered species, which often exhibit similarities with closely related species. To address this issue, this study proposes a method for classifying images of 11 parrot species using a hierarchical structure based on a CNN image classification model. The classification criteria in the first layer are defined by dividing the data based on the genus Cacatua and employing a random division approach. Experimental results comparing the proposed method with existing image classification techniques reveal that the precision improved from 0.917 to 0.952, the recall improved from 0.914 to 0.950, and the F1 score increased by 0.036, from 0.914 to 0.950. Additionally, when the images were randomly sorted in the first layer, the precision improved from 0.917 to 0.944, the recall improved from 0.914 to 0.943, and the F1 score increased by 0.029, from 0.914 to 0.943. These results indicate that the first-stage classification based on the biological taxonomic system led to an improvement of approximately 20% in terms of the F1 score compared to when the first-stage classification was done randomly in groups. Consequently, the proposed method demonstrates better performance in animal image classification problems when initially pre-classifying visually similar classes. On the other hand, if the external appearances of all classes are entirely different, the proposed method may not be suitable for application.