Hadamard Coding for Supervised Discrete Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Binary hashing is crucial for efficient large-scale data retrieval (images, videos, documents).
• Supervised Discrete Hashing (SDH) methods efficiently solve complex optimization problems for binary code generation.
• Existing methods like SDH and Iterative Quantization (ITQ) can be computationally intensive and sensitive to initial values.

Purpose of the Study:

• To propose a simplified and more efficient learning-based supervised discrete hashing method.
• To leverage Hadamard matrices for an exact mathematical solution in binary hashing.
• To improve upon the performance and computational efficiency of existing SDH techniques.

Main Methods:

• Developed Hadamard coded supervised discrete hashing (HC-SDH), a novel learning-based hashing approach.
• Derived a mathematically exact solution using Hadamard matrices, eliminating the need for alternating optimization.
• Simplified the SDH model without compromising performance.

Main Results:

• HC-SDH demonstrates superior precision, recall, and computational speed compared to conventional SDH.
• Achieved higher mean average precision (mAP) and top-accuracy on large datasets (SUN-397, ImageNet) than SDH and FastHash.
• HC-SDH training is 170x faster than SDH, and testing is 7x faster than FastHash.

Conclusions:

• HC-SDH offers a significant advancement in supervised discrete hashing, providing a faster and more accurate solution.
• The method's simplicity, lack of reliance on initial values, and superior performance make it highly suitable for large-scale retrieval tasks.
• Hadamard coding provides an effective and efficient approach to generating exact binary codes for supervised hashing.