Coordinate feature fusion networks for fine-grained image classification

Learning feature representations from discriminative local features plays a key role in fine-grained classification, but many methods tend to focus only on salient features in images and ignore most latent features. Exploiting rich features between channels and spaces helps capture this difference. Based on this idea, this paper proposes a Coordinate Feature Fusion Network (CFFN), which can be modeled by the channel and spatial feature interactions of images. CFFN consists of Feature Enhancement and Suppression Modules (FESM) and Coordinate Feature Interaction Module (CFIM). FESM gets saliency factors by aggregating the most salient parts in spatial and channel features, which obtains salient features through feature mapping, and suppresses the obtained salient features to force the network to mine the remaining latent features. Through the saliency and latent feature modules, more discriminative features can be effectively captured. The CFIM module can explore feature correlations in images, and the model learns complementary features from related channels and spaces, resulting in stronger fine-grained features. Our model can be trained in an end-to-end manner and does not require bounding boxes. It achieves 89.5, 93.4 and 94.8% accuracy on three benchmark datasets CUB-200-2011, FGVC-Aircraft and Stanford Cars, respectively.