Class-specific Reconstruction Transfer Learning for Visual Recognition Across Domains.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Subspace learning and reconstruction are key in transfer learning for knowledge sharing across domains.
• Existing domain adaptation methods often neglect class priors, leading to biased transfer functions, especially with limited data per class.

Purpose of the Study:

• To propose a novel class-wise reconstruction-based adaptation method, Class-specific Reconstruction Transfer Learning (CRTL).
• To address the limitations of existing methods by fully exploiting intra-class dependency and inter-class independency.

Main Methods:

• CRTL utilizes a class-specific reconstruction matrix to align source and target domains, leveraging class priors for domain distribution consistency.
• Introduces a projected Hilbert-Schmidt Independence Criterion (pHSIC) to maintain data-label relationships after feature augmentation.
• Applies low-rank and sparse constraints on the reconstruction coefficient matrix to preserve domain correlation structures.

Main Results:

• CRTL demonstrates superior performance over state-of-the-art representation-based domain adaptation methods on benchmark datasets.
• The class-specific approach effectively mitigates bias caused by data scarcity in specific classes.

Conclusions:

• CRTL offers a robust solution for domain adaptation by incorporating class-specific information and preserving data structure.
• The method enhances cross-domain classification accuracy, particularly in scenarios with imbalanced class distributions.