Efficient Semi-Supervised Multimodal Hashing With Importance Differentiation Regression.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Multi-modal hashing methods aim to create compact binary hash codes from heterogeneous features for large-scale multimedia retrieval.
• Supervised hashing requires extensive labeled data, while unsupervised methods struggle with semantic correlations.
• Existing methods face limitations due to reliance on labeled data or inability to capture multi-modal semantic relationships.

Purpose of the Study:

• To propose an Efficient Semi-supervised Multi-modal Hashing with Importance Differentiation Regression (ESMH-IDR) model.
• To address the limitations of supervised and unsupervised hashing by effectively utilizing both labeled and unlabeled data.
• To enhance the performance and efficiency of multi-modal hashing for multimedia retrieval.

Main Methods:

• Developed an efficient semi-supervised multi-modal hash code learning module using asymmetric learning for labeled data.
• Employed nonlinear regression with a shared projection matrix to preserve the structure of unlabeled data.
• Introduced an importance differentiation regression strategy to weigh hash codes from labeled and unlabeled samples differently.
• Utilized an efficient discrete optimization method with guaranteed convergence for hash optimization.

Main Results:

• The proposed ESMH-IDR model demonstrates superior retrieval effectiveness compared to existing methods.
• The method achieves significant improvements in retrieval efficiency.
• Experiments on public multimedia retrieval datasets validate the model's performance.

Conclusions:

• ESMH-IDR effectively alleviates the problems of supervised and unsupervised hashing by integrating labeled and unlabeled data.
• The importance differentiation strategy enhances the learning of hash functions.
• The model offers a promising solution for large-scale multimedia retrieval applications.