Fine-Grained Image Retrieval via Object Localization

In this paper, a network consisting of an object localization module and discriminative feature extraction is designed for fine-grained image retrieval (FGIR). order to reduce the interference complex backgrounds, introduced into before extraction. By selecting convolutional descriptors, main separated from background, thus, most filtered out. Further, in improve overall performance network, filter bank as local detector. Hence, features can be extracted directly original map. The experimental results based on CUB-200-2011 Cars-196 datasets demonstrate that proposed method FGIR.