A Multi-Grained Attention Residual Network for Image Classification

Attention mechanisms in deep learning can focus on critical features and ignore irrelevant details in the target task. This paper proposes a new multi-grained attention model (MGAN) to extract parts from images. The model includes a multi-grain spatial attention (MSA) mechanism and a multi-grain channel attention (MCA) mechanism. We use different convolutional branches and pooling layers to focus on the crucial information in the sample feature space and extract richer multi-grain features from the image. The model uses ResNet and Res2Net as the backbone networks to implement the image classification task. Experiments on the CIFAR10/100 and Mini-Imagenet datasets show that the proposed model MGAN can better focus on the critical information in the sample feature space, extract richer multi-grain features from the images, and significantly improve the image classification accuracy of the network.