Context-aware attention network for image recognition

Existing recognition methods based on deep learning have achieved impressive performance. However, most of these algorithms do not fully utilize the contexts and discriminative parts, which limit the recognition performance. In this paper, we propose a context-aware attention network that imitates the human visual attention mechanism. The proposed network mainly consists of a context learning module and an attention transfer module. Firstly, we design the context learning module that carries on contextual information transmission along four directions: left, right, top and down to capture valuable contexts. Second, the attention transfer module is proposed to generate attention maps that contain different attention regions, benefiting for extracting discriminative features. Specially, the attention maps are generated through multiple glimpses. In each glimpse, we generate the corresponding attention map and apply it to the next glimpse. This means that our attention is shifting constantly, and the shift is not random but is closely related to the last attention. Finally, we consider all located attention regions to achieve accurate image recognition. Experimental results show that our method achieves state-of-the-art performance with 97.68% accuracy, 82.42% accuracy, 80.32% accuracy and 86.12% accuracy on CIFAR-10, CIFAR-100, Caltech-256 and CUB-200, respectively.