Deep Residual Correction Network for Partial Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep domain adaptation (DDA) methods excel at transferring knowledge from labeled source to unlabeled target domains.
• Existing DDA often assumes identical label spaces, limiting real-world applicability.
• Partial domain adaptation (PDA) addresses scenarios where the target label space is a subset of the source label space.

Purpose of the Study:

• To propose an efficient Deep Residual Correction Network (DRCN) for partial domain adaptation.
• To enhance feature representation and mitigate negative impacts from irrelevant source classes in PDA.
• To improve cross-domain visual recognition performance under partial label set assumptions.

Main Methods:

• Introduced DRCN by integrating a residual block and task-specific feature layer into the source network.
• Employed fully-connected layers within the residual block to deepen the network and enhance feature representation.
• Designed a weighted class-wise domain alignment loss to match feature distributions of shared classes.

Main Results:

• DRCN demonstrated superior performance compared to existing deep domain adaptation approaches.
• The proposed method effectively enhanced adaptation from source to target domains.
• Experiments confirmed DRCN's effectiveness on partial, traditional, and fine-grained cross-domain visual recognition tasks.

Conclusions:

• DRCN offers an efficient and effective solution for partial domain adaptation challenges.
• The network architecture and alignment loss successfully address the complexities of subset label spaces.
• DRCN advances the state-of-the-art in cross-domain visual recognition.