Supervised Discrete Hashing With Relaxation.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• High-dimensional data retrieval and storage necessitate efficient methods.
• Data-dependent hashing techniques are gaining traction for their effectiveness.
• Existing supervised discrete hashing (SDH) methods use fixed regression targets.

Purpose of the Study:

• To propose a novel learning-based hashing method, supervised discrete hashing with relaxation (SDHR).
• To enhance the accuracy and flexibility of hashing by optimizing regression targets.
• To demonstrate the superiority of SDHR over traditional SDH methods.

Main Methods:

• Developed SDHR, a learning-based hashing method building upon SDH.
• Optimized regression targets within SDHR, incorporating a large margin constraint for classification.
• Employed ordinary least squares regression with a learned regression target matrix.

Main Results:

• SDHR demonstrated superior performance compared to the standard SDH method.
• Experimental results on CIFAR-10, MNIST, and FRGC datasets validated SDHR's effectiveness.
• The optimized regression target matrix in SDHR provided more accurate classification error measurement.

Conclusions:

• SDHR offers a more flexible and accurate approach to supervised discrete hashing.
• The method significantly improves retrieval and storage efficiency for high-dimensional data.
• SDHR represents an advancement in learning-based hashing for large-scale datasets.