Fine-grained image classification using deep learning

Abstract

Fine-grained image categorization, also known as sub-category recognition, is a popular research topic in computer vision and pattern recognition in recent years. The goal of this task is to classify images belonging to the same basic category (e.g., car, dog, flower, bird, etc.) into more detailed sub-categories. Compared with ordinary image classification tasks, fine-grained image categorization is a more challenging task due to the subtle inter-class differences and large intra-class variations between sub-categories.

In order to successfully classify two very similar species at a fine-grained level, it is most important to find the discriminative part in the images that can distinguish the two species, and to be able to represent the characteristics of these discriminative parts well. Most of existing methods for fine-grained classification of images are based on deep convolutional networks to learn a robust representation of image features. However, it is difficult to accomplish fine-grained image classification tasks with high accuracy by relying on a backbone neural network alone.

This dissertation aims to evaluate the performance of two different network-based fine-grained classification models, i.e., Progressive Multi-Granularity training framework and A Transformer Architecture for Fine-grained Recognition, on a wide range of the special-purpose fine-grained classification datasets. Both methods are currently state-of-the-art. The detailed and comprehensive experiments and in-depth analysis of the results provide valuable insight for future study on fine-grained image classification.