Researches advanced in fine-grained image classification based on convolutional neural network

Abstract

Due to practical needs, fine-grained image classification (FGIC) has been considered for many years as a direction of study in computer vision, which aims to subdivide images belonging to one coarse-grained category into multiple fine-grained classes. Traditional fine-grained image classification algorithms rely heavily on annotations. Recently, convolutional neural networks (CNN) have prefigured unprecedented opportunities for this research direction with the popularity and development in deep learning. To start, this study introduces the development history with various fine-grained image classification algorithms, as well as definition and research significance of the problem. After that, it compares and analyzes the different algorithms respectively in the aspects of strong supervision and weak supervision. This paper also compares the accuracy of these models on frequently used datasets. We conclude the paper by summarizing and evaluating the different aspects of these algorithms, and then discuss possible future research domains and challenges in this field.