Feature dispersion degree-based attention mechanism method and system

Abstract

The invention relates to the technical field of image processing, in particular to a feature dispersion degree-based attention mechanism method and system applied to a convolutional neural network, and the method comprises the steps: obtaining a target variance corresponding to a pixel value in an initial feature map based on the initial feature map; based on the target variance and a sigmoid activation function, obtaining a target weight value; and obtaining a target feature map based on the target weight value and the initial feature map. According to the invention, the weight distribution of the neural network to the feature map can be optimized, so that the performance of the neural network in classification and detection is improved.

Description

technical field [0001] The present application relates to the technical field of image processing, and in particular, to a method and system for an attention mechanism based on the degree of discreteness of features applied to a convolutional neural network. Background technique [0002] At present, the weight acquisition method of the existing channel attention mechanism applied to convolutional neural networks usually uses global pooling (max pooling, average pooling or a combination of both) to obtain the mean or maximum value of each channel, so that Each channel has its own weight. Connect these weights to a fully connected layer to finetune the weights, and then use the sigmoid function to normalize them to correspond to the original feature map channels. Multiply to achieve channel weight redistribution to feature maps. [0003] However, this allocation method does not take into account the influence of the richness (discreteness) of the feature itself on the recogni...

AI Technical Summary

Problems solved by technology

[0003] However, this allocation method does not take into account the influence of the richness (discrete degree) of the feature itself on the effectiveness of recognition, that is, for example, max pooling is used to obtain the maximum value of each channel, but the channel corresponding to the maximum value does not It must have the most effective features for classification, or it may be that there is just a certain feature point in this feature map that has particularity

In the same way, such a method (considering the global maximum or average of the absolute value) sometimes weakens the importance of some high-rich features, such as a feature with a more complex feature shape, that is, the pixel value in the feature has The average pooling method will directly ignore the detail richness of this feature, thus losing the consideration of the importance of this feature and affecting the weight redistribution of the feature map.

Images