Studies Advanced in Weakly Supervised Fine-grained Image Classification based on Deep Learning

Abstract

Recently, the concept of fine-grained classification has arouse much attention., which has aroused heated disscussion of academia and industry. The extraction of picture characteristics in early efforts on fine-grained image classification relied on dense annotations, but acquiring these annotations was time-consuming and labor-intensive. Lately, weakly supervised fine-grained image classification has gradually emerged, which can mainly be separated between techniques using the attention mechanism and techniques using various neural networks. In this paper, focusing on the above two types of frameworks, we first introduce representative algorithms, including their innovation, basic processes, advantages and disadvantages. We then quantitatively compare the results of different algorithms on mainstream datasets, where the attention based methods can achieve excellent accuracy. We finally discuss the existing challenges and future development of the fine-grained classification task, which we believe can provide some new insight for this task.