Deep learning method for lightweight bottleneck attention mechanism

Abstract

The invention discloses a deep learning method for a lightweight bottleneck attention mechanism. The attention mechanism is divided into a channel attention branch and a space attention branch. The channel attention branch obtains space information of the middle feature map through global average pooling and maximum value pooling, information of k channels in a left neighborhood and a right neighborhood with the channel as the center is rolled up and combined, the two obtained feature maps are added according to elements, and channel attention is generated; the spatial attention branch uses convolution to perform dimension reduction, context acquisition is performed through two dilated convolutions, and finally, the convolution is used to compress the number of channels to one dimension to generate spatial attention; according to the method, the two branch attention are fused, channel attention is broadcasted and expanded according to two dimensions of space, then space attention is broadcasted and expanded according to channel dimensions, the two expanded attention branches are added according to elements, and lightweight bottleneck attention is generated through Sigmoid function operation. The method is low in calculation cost and high in model learning ability.

Description

technical field [0001] The invention relates to the technical field of attention mechanism learning, in particular to a deep learning method of a lightweight bottleneck attention mechanism. Background technique [0002] In recent years, Convolutional Neural Networks (CNN) has rapidly promoted the development of computer vision, and it has demonstrated powerful performance in tasks such as image classification, target detection, and semantic segmentation. In order to further enhance the feature expression ability of CNN, recent research mainly focuses on three important factors of the network model: network depth, network width and network cardinality. [0003] In addition to the above three factors, many researchers have integrated the attention module into the convolution module in recent years, which proves that the attention module has great potential in improving the feature expression ability of the network. [0004] The attention module improves the performance of CNN...

AI Technical Summary

Problems solved by technology

[0005] In summary, the existing attention mechanism will greatly increase the overhead of the model when it is integrated into other network models. At present, it is urgent to reduce the overhead of the attention mechanism.

Images