Fast Supervised Discrete Hashing.

Area of Science:

• Machine Learning
• Computer Vision
• Data Mining

Background:

• Learning-based hashing methods are crucial for scaling data retrieval.
• Traditional supervised discrete hashing (SDH) involves time-consuming iterative optimization.

Purpose of the Study:

• To propose a novel, efficient learning-based hashing algorithm named fast supervised discrete hashing (FSDH).
• To accelerate the hashing process while maintaining or improving accuracy compared to existing methods.

Main Methods:

• FSDH employs a unique regression strategy, mapping class labels to hash codes, differing from conventional methods.
• It features a closed-form solution for hash code optimization, eliminating iterative steps.
• The method efficiently solves the projection matrix for least squares regression.

Main Results:

• FSDH demonstrates significant speed improvements, e.g., 12x faster than SDH on CIFAR-10 (128 bits) and 151x faster than FastHash on MNIST (64 bits).
• Experimental results indicate FSDH outperforms comparative methods in both speed and accuracy.

Conclusions:

• FSDH offers a highly efficient and effective solution for learning-based hashing.
• The proposed regression strategy and closed-form solution provide substantial computational advantages.