Progressive learning for weakly supervised fine-grained classification

Tiantian Yan, Shijie Wang, Zhihui Wang, Haojie Li, Zhongxuan Luo

Abstract: Highlights • We propose progressive patch localization module to solve the problem that the selected patches with lower rank very likely contain noise information while guaranteeing a diversity of fine-grained features. • Feature calibration module is proposed to calibrate patch-level features for strengthening its discriminative information and suppressing useless information by employing the global information, which further benefits the final classification performance. • We evaluate our method on three challenging datasets (CUB, Cars and Aircraft), and achieve the state-of-the-art results on all of these datasets. Abstract Despite fine-grained image classification has made considerable progress, it still remains a challenging task due to the difficulty of finding subtle distinctions. Most existing methods solve this problem by selecting the top-N highest scores’ discriminative patches from candidate patches at one time. However, since the classification network often highlights small and sparse regions, the selected patches with the lower rank may contain noise information. To address this problem and ensure the diversity of fine-grained features, we propose a progressive patch localization module (PPL) to find the discriminative patches more accurately. Specifically, this work employs the classification model to find first most discriminative patch, then removes the most salient region to help the localization of the next most discriminative patch, and the top-K discriminative patches can be found by repeating this procedure. In addition, in order to further improve the representational power of patch-level features, we propose a feature calibration module (FCM). This module employs the global information to selectively emphasize discriminative features and suppress useless information, which can obtain more robust and discriminative local feature representations and then help classification network achieve better performance. Extensive experiments are conducted to show the substantial improvements of our method on three benchmark datasets.