Fine-grained visual classification via multilayer bilinear pooling with object localization

Fine-grained visual classification is a challenging task in the computer vision field. How to explore discriminative features is vital for classification. As one crucial step, exactly object localization is able to eliminate the background noises and highlight interesting objects at the same time. However, some current methods usually use bounding boxes to locate objects, that are not suitable when the poses of objects change. Furthermore, it has been demonstrated that deep features have strong feature representation capability, especially the bilinear pooling features, which achieved superior performance in fine-grained visual classification tasks. However, the bilinear features, which captured only from the last convolutional layer, have limited discriminability, especially when dealing with small-scale objects. In this paper, we propose a multilayer bilinear pooling model combined with object localization. First, a flexible and scalable object localization module is utilized to locate the interesting object in an image instead of using bounding boxes. Then the refined features are obtained by highlighting object region and suppressing background noises. While the multilayer bilinear pooling, which exploits the complementarity between different layers, is used for further extracting more discriminative features. Experiment results on three public datasets show that our proposed method can achieve competitive performance compared with several state-of-the-art methods.