An Efficient Regenerated Cross-Modal Hashing: Improving Existing Hash Codes with the Arbitrary Length.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Cross-modal retrieval relies on hashing techniques for efficiency.
• Fixed hash code lengths in deployed models limit adaptability to diverse retrieval scenarios.
• Retraining models for different hash code lengths is computationally expensive and may degrade performance.

Purpose of the Study:

• To introduce Regenerated Cross-Modal Hashing (RCMH), a novel framework for improving existing hash codes and adapting them to arbitrary lengths.
• To address the challenges of retraining models for different hash code lengths in cross-modal retrieval.
• To enhance retrieval accuracy and efficiency by optimizing hash code quality and length flexibility.

Main Methods:

• Initialize target-length hash codes by clipping/padding existing codes, guided by an augmented label similarity matrix.
• Employ a linear-nonlinear competitive reconstruction approach to bridge semantic gaps between image and text features.
• Utilize a regenerate hashing term for efficient final hash code generation without iterative optimization, preserving information from existing codes.

Main Results:

• RCMH successfully converts existing hash codes to arbitrary lengths with high efficiency.
• The method reduces semantic gaps and captures deeper modality-specific relationships.
• RCMH integrates with state-of-the-art methods, improving their hash code length adaptability and retrieval performance.
• Experiments demonstrate RCMH outperforms existing methods on public datasets across various metrics.

Conclusions:

• RCMH offers an efficient and effective solution for flexible hash code length adjustment in cross-modal retrieval.
• The framework enhances the quality of existing hash codes, leading to improved retrieval accuracy.
• RCMH provides a robust and computationally efficient alternative to model retraining for adapting hash code lengths.