Discriminative Transfer Feature and Label Consistency for Cross-Domain Image Classification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Visual domain adaptation (VDA) aims to transfer knowledge from labeled source domains to unlabeled target domains with different data distributions.
• Existing VDA methods often focus on domain-invariant feature extraction but overlook category discriminability and the impact of noisy pseudo-labels.
• Misclassified target pseudo-labels can negatively affect cross-domain alignment and hinder model performance.

Purpose of the Study:

• To propose a novel approach, Discriminative Transfer Feature and Label Consistency (DTLC), to address limitations in current visual domain adaptation techniques.
• To simultaneously ensure learned features are both domain-invariant and category-discriminative.
• To mitigate the negative effects of inaccurate target pseudo-labels on cross-domain alignment.

Main Methods:

• DTLC unifies cross-domain alignment with discriminative information preservation and label consistency within a single framework.
• It incorporates class discriminative information by adjusting feature distances within and between classes during distribution alignment.
• Target pseudo-labels are refined iteratively based on label consistency, creating a coupled learning process with feature transfer.

Main Results:

• DTLC demonstrated significant improvements over state-of-the-art non-deep visual domain adaptation methods.
• The proposed method achieved performance comparable to competitive deep domain adaptation approaches.
• Experiments on multiple visual cross-domain benchmarks validated the effectiveness of the DTLC approach.

Conclusions:

• The DTLC approach effectively addresses the dual challenges of learning discriminative, domain-invariant features and handling noisy pseudo-labels in VDA.
• The iterative coupling of feature learning and label refinement leads to superior cross-domain adaptation performance.
• DTLC offers a promising direction for advancing visual domain adaptation research.