A Multi-source Domain Adaptive Model and Method Based on Partial Feature Alignment

Abstract

The invention discloses a multi-source domain adaptive model and method based on partial feature alignment, wherein the feature selection module of partial feature extraction is based on the conventional convolutional neural network or residual neural network feature extractor, according to the source domain and The similarity of each feature dimension of the target domain generates a selection vector at the feature level. After the selection vector is applied to the initial feature map, some features in the source domain that are highly related to the target domain can be screened out. On this basis, the present invention further proposes three partial feature alignment loss functions for intra-category, inter-domain and inter-category, so that the purified feature map is more distinguishable for the classifier, and the source domain and target domain Relevant partial features are highlighted. The invention is used for multi-source domain self-adaptive classification data sets, and compared with the existing multi-source domain self-adaptive model, the classification accuracy rate is higher and the effect of feature selection is better.

Description

technical field [0001] The invention belongs to the multi-source domain self-adaptive branch in computer vision and migration learning, and specifically relates to a multi-source domain self-adaptive model structure that aligns partial features through high-order moment matching, and designs an alignment loss function on the feature map to To achieve the purpose of distinguishing features from classifiers. Background technique [0002] In machine learning, supervised and semi-supervised learning using deep neural networks has achieved remarkable results. Relying on many public data sets, supervised and semi-supervised learning can be used in various tasks such as image classification, face recognition, and semantic analysis. Has a wide range of applications. However, the collection of data labels in the real world is very difficult and often requires a lot of manpower to complete. Due to the existence of inter-domain offset, the models trained on other data sets cannot be d...

AI Technical Summary

Problems solved by technology

For example, when performing feature alignment, some existing methods use all image features obtained through feature extractors, which not only include features related to the source domain and target domain, but also include The domain-specific features that appear in the source domain, if these features also participate in the feature alignment, will cause a decline in the alignment effect

Images