Attention weight module and method for convolutional neural network

Abstract

The invention discloses an attention weight module and method for a convolutional neural network. The attention weight module comprises: an extraction unit configured to extract a feature vector from a feature map input to a convolutional neural network; a generation unit configured to feed the feature vector to the fully connected layer to generate an attention vector; and a weight allocation unit configured to allocate a weight to the feature map based on the attention vector. The attention weight block can extract more semantic information by considering the spatial attention weight and the channel relation weight at the same time, so that the convolutional neural network can have higher expression ability.

Description

technical field [0001] The invention relates to the field of image processing. Specifically, the present invention relates to attention weight modules and methods. Background technique [0002] In recent years, deep neural networks (deep learning) have been widely used in the field of pattern recognition, and many breakthroughs have been achieved, for example, using deep neural networks to achieve object detection, face recognition, semantic segmentation, etc. The achievement of these achievements is closely related to the ability of deep neural network to learn features with strong expressive ability (image representation method). Compared with the traditional manual feature extraction method, the manual method extracts the low-level features of the image by introducing expert knowledge, while the deep neural network is driven by data and can learn the hierarchical features of the image, including rich low-level features. (low-level), mid-level (mid-level), and high-level...

AI Technical Summary

Problems solved by technology

The disadvantages of this technique are: (1) RAN ignores the importance of the relationship between channels and treats each feature map equally

[0011] SE-Net does not effectively process feature maps in the spatial dimension. Human vision research shows that human vision can quickly notice key areas, while SE-Net does not process in the spatial domain, and is not robust to image spatial information transformation. Rod

Images