Multiple attention relation network for few-shot learning

In a large amount of practical application scenarios, only small number of labeled samples could be collected for classification model training, which makes the big dataset based deep networks inapplicable. Few-shot learning methods have been developed to solve the small sample problem with only a few samples for each category. Among which, relation network has been an efficient method based on convolutional modules to compute the sample relations. Due to the local receptive field of convolutional operation, it could not obtain features on a global level and ignores the contextual structures within the samples. The discriminative salient features are not emphasized either. To address these issues, a multiple attention relation network (MARN) is proposed which combined spatial attention, channel attention, self attention and cross attention. Spatial attention and channel attention can obtain global description in space and across channels. These two attention module selectively aggregates the features from separate parts by weighted sum. Self attention and cross attention can capture the self and cross relations between deep features, which could enhance the salient features and increase the model comparison ability. Besides these attention mechanisms, atrous spatial pyramid pooling (ASPP) is employed to extract features from different scale receptive fields. Accordingly, MARN can efficiently combine local and global spatial features, intra and inter channel features, and different scale features, which can improve the model discrimination capability and classification performance. Meanwhile, the combination of these mechanisms only introduce a very small number of parameters. Experiments on few shot learning benchmarks mini-ImageNet and CUB-200 have verified the superiority of MARN over the state-of-the-art methods.