Image classification method, system and electronic device based on dual attention mechanism

Abstract

The invention discloses an image classification method, system and electronic equipment based on a dual attention mechanism. Using an improved self-attention model, the attention weights of the feature map space dimension and the feature map channel dimension are simultaneously calculated, and then the two kinds of attention The feature maps generated by the force mechanism are added together to complete the end-to-end image classification network construction, improve the network's adaptive perception of salient areas, achieve more accurate and robust image feature expression, and then improve image classification performance.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to an image classification method, system and electronic equipment, in particular to an image classification method, system and electronic equipment based on a dual attention mechanism. Background technique [0002] Since the birth of deep learning, related technologies, especially those related to deep convolutional networks, have brought great changes and influences to the field of computer vision. The image classification method based on deep convolutional networks has become the mainstream method of current research. However, this method also has the following disadvantages: [0003] (1) Excessive reliance on training data. Image feature expression is the core content of image classification methods based on deep convolutional networks, and accurate and robust feature acquisition largely depends on the network's learning of massive data. The quantity and quality of data ...

AI Technical Summary

Problems solved by technology

[0004] (2) Limitations of convolution mode

The essence of the convolution operation is to use a single sampling mode to filter a specific area. Although the deep convolutional network can realize the feature expression in the large-scale receptive field through multi-layer and multi-modal convolution operations, the singleness and adjacency of the sampling mode This limitation restricts the accurate expression of features of different scales by the convolutional network, and also leads to the weak correlation of the local representation of the final generated features at the overall image level, which cannot ensure that the local features with similar semantic content of the image have similar feature expression

[0005] (3) The classification recognition degree between the feature map channels is different

Images