Fast Fine-Grained Image Classification via Weakly Supervised Discriminative Localization

Abstract

Fine-grained image classification is to recognize hundreds of subcategories\nin each basic-level category. Existing methods employ discriminative\nlocalization to find the key distinctions among subcategories. However, they\ngenerally have two limitations: (1) Discriminative localization relies on\nregion proposal methods to hypothesize the locations of discriminative regions,\nwhich are time-consuming. (2) The training of discriminative localization\ndepends on object or part annotations, which are heavily labor-consuming. It is\nhighly challenging to address the two key limitations simultaneously, and\nexisting methods only focus on one of them. Therefore, we propose a weakly\nsupervised discriminative localization approach (WSDL) for fast fine-grained\nimage classification to address the two limitations at the same time, and its\nmain advantages are: (1) n-pathway end-to-end discriminative localization\nnetwork is designed to improve classification speed, which simultaneously\nlocalizes multiple different discriminative regions for one image to boost\nclassification accuracy, and shares full-image convolutional features generated\nby region proposal network to accelerate the process of generating region\nproposals as well as reduce the computation of convolutional operation. (2)\nMulti-level attention guided localization learning is proposed to localize\ndiscriminative regions with different focuses automatically, without using\nobject and part annotations, avoiding the labor consumption. Different level\nattentions focus on different characteristics of the image, which are\ncomplementary and boost the classification accuracy. Both are jointly employed\nto simultaneously improve classification speed and eliminate dependence on\nobject and part annotations. Compared with state-of-the-art methods on 2\nwidely-used fine-grained image classification datasets, our WSDL approach\nachieves the best performance.\n