Deep Learning for Fine-Grained Image Recognition: A Comprehensive Study

Abstract

In computer vision, image recognition is a noteworthy and hot research area which develops rapidly. The principal task of this technique is to automatically predict which pre-defined categories an image might belong to. Traditional image recognition targets to classify images into diversified highly distinguished categories. However, Fine-Grained Image Recognition (FGIR) aims to recognize the variances among images categorized in subordinate classes, e.g., species of birds, types of cars or species of flowers, which are equivalent to "species" in Taxonomy in certain aspects. As a result, models of FGIR are required to pick out features from finer granularity. Conventional methods apply special feature encoding to explore discernible attributes, while recent methods of FGIR makes great advancement with assistance of deep learning which has obtained the remarkable development nowadays. In this paper, we provide a new integration of the current leading FGIR models according to how they improve the development of FGIR. We classified them into five main categories and then compared their performance on three popular datasets and analyzed the results. To advance the further development of this topic, we point out some open problems worth further exploring.