Symmetrical bidirectional knowledge alignment for zero-shot sketch-based image retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot sketch-based image retrieval (ZS-SBIR) faces challenges due to significant cross-modality discrepancies.
• Existing methods often lack efficient bidirectional knowledge alignment between teacher and student models, limiting generalization.

Purpose of the Study:

• To propose a novel Symmetrical Bidirectional Knowledge Alignment (SBKA) framework for ZS-SBIR.
• To enhance the generalizability of ZS-SBIR models by effectively aligning knowledge between teacher and student models.

Main Methods:

• Developed a symmetrical bidirectional knowledge alignment learning framework (SBKA).
• Implemented a one-to-many cluster cross-modality matching method to leverage intra-class relationships and mitigate modality gaps.

Main Results:

• The proposed SBKA algorithm achieved superior performance on representative ZS-SBIR datasets (Sketchy Ext, TU-Berlin Ext, QuickDraw Ext).
• Demonstrated effectiveness in reducing the adverse effects of the modality gap in ZS-SBIR tasks.

Conclusions:

• SBKA framework effectively learns mutual discriminative information for improved knowledge alignment.
• The novel matching strategy enhances ZS-SBIR performance, outperforming state-of-the-art methods.