Deep Class-Wise Hashing: Semantics-Preserving Hashing via Class-Wise Loss

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Deep supervised hashing is effective for large-scale semantic image retrieval.
• Current methods using pairwise/triplet labels struggle with complex semantic concepts and have high training costs.

Purpose of the Study:

• To propose a novel deep supervised hashing model for learning compact, class-level similarity-preserving binary codes.
• To improve retrieval performance by directly using semantic labels for supervised information.

Main Methods:

• Developed a deep metric learning-based model using semantic labels for training.
• Introduced a novel cubic constraint loss function based on Gaussian distribution to preserve semantic variations and penalize class overlaps.
• Implemented a two-step optimization strategy to address discrete optimization and gradient vanishing issues.

Main Results:

• The proposed model learns more compact and discriminative hashing codes.
• Achieved state-of-the-art retrieval performance on five large-scale benchmark databases.
• The cubic constraint loss effectively preserves semantic variations and separates different classes in the embedding space.

Conclusions:

• The novel deep supervised hashing model significantly enhances semantic image retrieval.
• Directly utilizing semantic labels and the proposed loss function offers a more effective approach than traditional methods.
• The model demonstrates superior performance and efficiency in large-scale image retrieval tasks.