Attention-based CNN category activation graph generation method

Abstract

The invention belongs to the technical field of deep learning and computer visualization, and discloses an attention-based CNN category activation graph generation method, which comprises the steps of1, calculating a gradient value of each pixel point of a feature graph M = (M0, M1,..., MK-1) to serve as a spatial attention weight related to a neuron category; step 2, obtaining a connection weight corresponding to each type of neurons as a channel attention weight; and step 3, generating a CNN category activation graph according to the spatial attention weight and the channel attention weight. According to the method, the category activation weight acts on the attention weight, meanwhile, the channel-space position importance of the feature graph is utilized, compared with CAM and Grad-CAM methods, the provided method is better in visualization effect of the generated category activation graph, and the method is not limited by a network structure and is more flexible to use.

Description

technical field [0001] The invention belongs to the technical field of deep learning and computer visualization, in particular to a method for generating an attention-based CNN category activation map. Background technique [0002] Convolutional Neural Network (CNN) has achieved great success in many fields, but due to its end-to-end "black box" characteristics, it covers the knowledge storage and processing mechanism of the middle layer, making it impossible for people to spy on its internal characteristics and external features. The basis for decision-making affects its application value to a certain extent. Visualization is a common way to explain the reason of CNN's decision-making and display its internally learned features. At present, some researches apply it to CNN's feature understanding and decision-making reason explanation, such as the CAM method (class activation map) , Grad-CAM method (gradient-based CAM) (R.R.Selvaraju, M.Cogswell, A.Das, R.Vedantam, D.Parikh...

AI Technical Summary

Problems solved by technology

[0003] Although the image classification model based on the deep convolution network is getting higher and higher in accuracy, due to the limitation of the "end-to-end" attribute of the deep network, the classification process is like a "black box", and the classification result cannot be obtained. To explain, it is also impossible to locate the features of which areas of the image contribute the most to the classification results. An attention-based CNN category activation map generation method is proposed.

Images