It takes two: Dual Branch Augmentation Module for domain generalization.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) exhibit performance degradation on out-of-distribution data.
• Domain generalization (DG) aims to improve model robustness by learning from multiple source domains for unseen target domains.
• Existing Fourier transform-based DG methods inadequately address the domain gap by focusing only on suppressing source domain-specific information.

Purpose of the Study:

• To propose a novel Dual Branch Augmentation Module (DBAM) that effectively utilizes both amplitude and phase spectra for enhanced domain generalization.
• To address the limitations of prior methods by establishing connections between source and target domains and enhancing the utilization of domain-agnostic information.

Main Methods:

• DBAM leverages Fourier transform, utilizing both amplitude and phase spectra.
• The amplitude branch incorporates Inner-domain Amplitude Distribution Rectification (IADR) and Cross-domain Amplitude Dirichlet Mixup (CADM) for training stability and feature space exploration.
• The phase branch employs Random Symmetric Phase Perturbation (RSPP) to improve the robustness of domain-agnostic information recognition.

Main Results:

• DBAM significantly outperforms state-of-the-art (SOTA) methods in domain generalization tasks.
• Proposed techniques like Test-time Amplitude Prototype Calibration (TAPC) effectively mitigate the domain gap during evaluation.
• Extensive experiments on four benchmarks validate the effectiveness and superiority of the DBAM approach.

Conclusions:

• DBAM offers a significant advancement in domain generalization by effectively leveraging both amplitude and phase spectra.
• The proposed module enhances model robustness and generalization capabilities on unseen data.
• DBAM represents a promising direction for improving the reliability of deep neural networks in diverse environments.