Discrete Hashing with Multiple Supervision.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Supervised hashing methods leverage label information for accurate hash code generation.
• Existing methods often use large instance-pairwise similarity matrices, leading to high memory and computation costs.
• Relaxing discrete constraints in optimization can introduce quantization errors, degrading performance.

Purpose of the Study:

• To address the limitations of existing supervised hashing techniques.
• To propose a novel hashing method that is computationally efficient and minimizes quantization errors.
• To improve the accuracy of learned hash codes through multiple supervision signals.

Main Methods:

• Developed Discrete Hashing with Multiple Supervision (MSDH), a novel supervised hashing algorithm.
• MSDH utilizes both class-wise and instance-class similarity matrices, reducing space complexity compared to instance-pairwise methods.
• An iterative optimization algorithm is employed to directly learn discrete hash codes, avoiding constraint relaxation.

Main Results:

• MSDH demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.
• The proposed method achieves accurate and compact hash codes with reduced computational overhead.
• Directly learning discrete hash codes minimizes quantization errors, leading to better performance.

Conclusions:

• MSDH offers an effective and efficient solution for supervised hashing.
• The use of multiple supervision sources and direct discrete optimization enhances hash code quality.
• This approach is suitable for real-world applications demanding efficient and accurate hashing.