Map-and-acquisition networks

In recent years, attention mechanisms have greatly contributed to the representation ability of deep convolutional neural networks, especially the attention module in convolutional neural networks. This paper proposes an operational unit called Map-and-Acquisition (MA) to compute the attention of different feature vectors in convolutional neural networks based on simple mathematical principles. Inspired by the convolution operation, MA extends the convolution operator to integrate spatial or channel features in the local receptive field to the global spatial or channel, and only operates on one dimension of the image feature, avoiding a series of problems caused by the coupling of spatial and channel dimensions. The feature response of the channel or spatial domain is adaptively adjusted through the weighting and summing operator to better extract the high-level features of the image. Our method is verified through image classification, object detection, and instance segmentation on the ImageNet-1K and COCO 2017 datasets, and is also evaluated on the small-sample image classification task on the dataset mini-Imagenet. Experimental results show that our method can significantly improve the performance of vision tasks.