Feature Visualization Method for Deep Neural Networks Based on Constrained Optimization-like Activation Mapping

Abstract

The invention discloses a feature visualization method of a deep neural network based on constrained optimization class activation mapping. Obtain a pre-trained model built with a deep neural network for image classification by training or downloading; use the pre-trained model to forward pass an image to be tested to obtain a feature map, and further process to obtain the final weight vector; through the final weight The vector weights and sums the components of the feature map to obtain a visualized feature map, which is presented as the final visualization result. The present invention can perform feature visualization on any deep neural network, can achieve better visualization effect of deep feature interpretability, has less noise and stronger class discrimination.

Description

technical field [0001] The invention relates to an image feature visualization method in the field of deep learning interpretability, in particular to a deep neural network feature visualization method for constrained optimization and class activation mapping. Background technique [0002] Deep learning techniques have achieved remarkable results and superior performance in some fields, especially in the field of computer vision, such as image classification and other tasks. However, because its mathematical principles have not been fully proven, its end-to-end black-box nature makes it impossible for humans to know how a deep neural network makes decisions. Therefore, research on the interpretability of deep learning has gradually emerged in recent years. One of the most direct ideas is to use visualization technology to obtain image regions that play a positive role in prediction, especially to visualize the feature representation of the middle layer of deep neural network...

AI Technical Summary

Problems solved by technology

[0004] In order to solve the problems existing in the background, the present invention provides a deep neural network feature visualization method based on class activation mapping for the problem that the current deep neural network feature visualization results are noisy and the class discrimination is weak

Images