D-BIN: A Generalized Disentangling Batch Instance Normalization for Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-world pattern recognition faces challenges due to visual statistic variability.
• Existing algorithms struggle with generalization on unseen data due to the independent identically distributed assumption.
• Current domain adaptation methods lack interpretability.

Purpose of the Study:

• To address the poor generalization capability of pattern recognition algorithms.
• To investigate the role of style and content features in domain adaptation.
• To propose a novel, interpretable method for disentangling domain-specific and domain-invariant features.

Main Methods:

• Proposed a novel normalization module: disentangling batch instance normalization (D-BIN).
• Employed contrastive learning to align content representations and separate style representations across domains.
• Developed self-form and dual-form regularizers to preserve mutual information and match feature distributions.

Main Results:

• D-BIN effectively disentangles domain-specific and domain-invariant features.
• The proposed method improves performance in domain adaptation and generalization tasks.
• Experiments on various datasets demonstrate the effectiveness of D-BIN.

Conclusions:

• D-BIN offers an interpretable approach to enhance network generalization.
• The method successfully addresses the limitations of existing domain adaptation techniques.
• D-BIN can be readily integrated into state-of-the-art networks to boost performance.