Learning Binary Hash Codes Based on Adaptable Label Representations.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Supervised hashing aims to map images and semantic annotations to binary codes for efficient retrieval.
• Current deep supervised hashing methods often use suboptimal one-hot or multi-hot label encodings.
• These encodings fail to capture intricate semantic relationships between classes.

Purpose of the Study:

• To develop a novel deep supervised hashing approach that learns effective, data-driven class label representations.
• To improve the discriminative power of binary hash codes for image search.
• To address limitations of traditional label encoding methods in capturing semantic relationships.

Main Methods:

• Introduced Adaptive Labeling Deep Hashing (AdaLabelHash), a method that learns binary hash codes using adaptable class label representations.
• Treated class labels as trainable variables (codewords) on a K-dimensional hypercube, optimized during network training.
• Optimized hash mappings such that semantically similar images are represented by nearby codewords in Hamming distance.

Main Results:

• AdaLabelHash successfully learns label representations that effectively capture semantic relationships among classes.
• The approach yields compact and discriminative binary hash representations for images.
• Quantitative and qualitative evaluations demonstrated the superiority of AdaLabelHash over existing methods on benchmark datasets.

Conclusions:

• Learned label representations provide more discriminative signals for supervised hashing compared to traditional methods.
• AdaLabelHash offers an effective and easily implementable solution for learning both label representations and binary embeddings simultaneously.
• The proposed method significantly enhances the performance of binary codes for image retrieval tasks.