An Unsupervised Domain Adaptation Method Based on Adversarial Learning Loss Function

Abstract

The invention discloses an unsupervised domain adaptation method based on an adversarial learning loss function, including: (1) the source domain image generates high-level features through a feature extraction network G, and performs cross-entropy loss through a classifier C and real labels; on the other hand A confusion matrix is ​​generated by a domain discriminator D to correct pseudo-labels to real labels. (2) The target domain image generates high-level features through the feature extraction network G, and generates pseudo-labels through the classifier C. On the other hand, the high-level features generate a confusion matrix through the domain discriminator D, and correct the pseudo-labels to the opposite distribution. (3) Let the feature generator and the discriminator fight to optimize the above loss function; in addition, for the confusion matrix on the target domain, generate a correction label, and use it as the label of the target domain to optimize the classifier. The present invention can correct the noise of pseudo-labels and match the distribution differences between domains during unsupervised domain adaptation, thereby improving the classification accuracy of the target domain.

Description

technical field [0001] The invention belongs to the field of unsupervised domain adaptation of transfer learning, and in particular relates to an unsupervised domain adaptation method based on an adversarial learning loss function. Background technique [0002] In recent years, deep learning has made impressive progress on classification tasks. The success of deep neural networks is based on large-scale datasets with a large number of labeled samples. However, in many practical situations, a large number of labeled samples are not available. Deep neural networks pretrained on existing datasets do not generalize well to new data with different appearance features. Essentially, the difference in data distribution between domains makes it difficult to transfer knowledge from the source domain to the target domain. This transfer problem is called the domain transfer problem. [0003] Unsupervised domain adaptation aims to solve the above-mentioned domain transfer problem, in...

AI Technical Summary

Problems solved by technology

However, the alignment between the source and target feature distributions is implicit and not theoretically guaranteed

Without matching domain distribution, self-training-based methods lead to performance degradation in the case of shallow networks

Images