A Multi-scale Contextual Attention Mechanism For Convolutional Neural Networks

Abstract

In recent years, attention mechanism has been widely studied in the field of computer vision, which can effectively improve the performance of visual tasks. In the past, many classical attention models have studied the modeling of nonlinear relationships in the spatial or channel dimensions of feature maps, ignoring the use of contextual relationships to capture the information interaction of the three dimensions to obtain a global attention feature map. In this paper, we investigate an effective multi-scale contextual attention mechanism, which can obtain feature information of different receptive fields through the combination of multi-branch conventional convolution and dilated convolution, which can increase the image receptive field, and combine global features and detailed features to effectively use contextual information. In addition, since the input tensors interact with each other on the three dimensions of the feature map and was adjusted by an adaptive parameter, this also makes the three-dimensional attention weights we obtain more differentiated. Our MCA model is simple and can be flexibly embedded in a variety of classical backbone networks, and experimental evaluation of the proposed attention mechanism on common datasets for image classification and object detection also proves the effectiveness of our attention meachine.