Relaxed Asymmetric Deep Hashing Learning: Point-to-Angle Matching

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning excels in supervised hash function learning due to powerful data representation.
• Existing hashing methods often use overly strict point-to-point matching, limiting performance.

Purpose of the Study:

• To propose a novel deep supervised hashing method.
• To relax instance matching from point-to-point to point-to-angle.
• To improve similarity measurement for enhanced binary code generation.

Main Methods:

• Introduced an asymmetric inner product for similarity/dissimilarity measurement between real-valued output and binary codes.
• Developed a Hamming-distance-based triplet loss for ranking positive and negative pairs.
• Designed an algorithm for alternating optimization of deep features and binary codes.

Main Results:

• The proposed method relaxes output constraints, allowing wider choices beyond strict +1/-1.
• The asymmetric product projects outputs into the same Hamming space.
• Experiments on four datasets show the approach's effectiveness and superiority over state-of-the-art methods.

Conclusions:

• The novel point-to-angle matching strategy enhances deep supervised hashing.
• The asymmetric product and triplet loss effectively capture semantic relationships.
• The method offers a superior alternative to existing strict matching techniques in hashing.