Weakly Supervised Fine-Grained Image Recognition Based on Multi-Channel Attention and Object Localization

Abstract

The task of fine-grained image classification recognizes similar subcategories that belong to the same superclass. The subtle differences between inter classes and intra-class diversity make it a challenging task. The existing methods focus on how to locate the most discriminative parts, which usually ignored the whole feature of objects. In this paper, we propose a fine-grained image recognition method based on multi-channel attention and object localization, which mainly includes two parts: multi-channel attention (MCA) module is used to learn different discriminative regions, attention object location (AOL) module can locate the object from the input image. The method we proposed can be trained in an end-to-end manner without adding bounding box/part annotation. At the same time, we build a new application-oriented fine-grained dataset: Poyang Lake birds, which contain 370 species of birds in Poyang Lake. We conducted extensive experiments on our dataset and three commonly used fine-grained datasets (CUB-200-2011, Stanford Cars, and FGVC-Aircraft). Extensive experiments show that our model shows compelling performance whether it is on our dataset or the other existing datasets.