A Robust Feature Deep Learning Method Based on Adversarial Space Transformation Network

Abstract

The invention belongs to the field of artificial intelligence and discloses a robust feature deep learning method based on an adversarial space transformation network. Implementation steps: read the labeled image data in the training set; input the labeled image data into the classifier network and calculate the classifier loss; read the image data in the training set according to the requirements; input the read image data into the confrontation space transformation network Obtain the transformed image; input the pre-transformed and transformed images into the classifier network to obtain the output of the classifier corresponding to the original image and the transformed image; calculate the consistency of the classifier output values ​​​​corresponding to the original and transformed images Loss; weighted sum of classifier loss and consistency loss; optimize objective function, update adversarial space transformation network and classifier network parameters. The invention can effectively improve the generalization ability and self-adaptive ability of the deep learning model for samples with different distributions, and has better anti-interference ability for interference and defaced samples.

Description

technical field [0001] The invention belongs to the field of artificial intelligence machine learning, and in particular relates to a deep learning method for feature representation with domain invariance and robustness for tasks such as image classification and recognition. Background technique [0002] The training data of machine learning, especially deep learning models, usually comes from samples of limited distribution. When the model has to deal with data from different distributions, it will lead to a significant drop in performance. In addition, when the input data is disturbed or defaced, the accuracy and stability of deep learning models are also easily severely affected. [0003] Currently, these two issues are usually treated as separate issues. For the above-mentioned problem of model performance degradation caused by differences in data distribution, technical methods such as domain adaptation and domain generalization are usually used to solve the problem. ...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to solve the following technical problems, that is, when the training data comes from limited distribution samples and is affected by interference or defaced samples when inputting data, the existing deep learning model cannot adapt to solve the above two problems at the same time

Images