Local Semantic-aware Deep Hashing with Hamming-isometric Quantization.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Deep hashing methods leverage deep neural networks for powerful feature representation and hash function learning, outperforming traditional handcrafted feature approaches.
• Existing deep hashing methods often focus on pairwise or triplet-wise constraints, potentially neglecting crucial local data structures.
• This limitation can hinder the full effectiveness of hash learning for tasks like image retrieval and data storage.

Purpose of the Study:

• To propose a novel deep hashing method, Local Semantic-aware Deep Hashing (LSDH), that integrates local data similarity into the hash learning process.
• To address the limitation of existing methods by better exploiting local data structures for improved hash learning.
• To enhance the quality of binary codes through a novel Hamming-isometric quantization objective.

Main Methods:

• Developed LSDH, a deep hashing approach that explicitly incorporates local data similarity into hash learning.
• Utilized semantic relations within the Hamming space to robustly preserve local data similarity.
• Introduced a Hamming-isometric objective to maximize similarity consistency between binary-like features and their corresponding binary codes, reducing quantization error.

Main Results:

• LSDH demonstrated superior performance compared to state-of-the-art hashing methods on benchmark datasets.
• Experiments were conducted on single-label datasets (CIFAR-10, CIFAR-20, SUN397) and a multi-label dataset (NUS-WIDE).
• The proposed Hamming-isometric quantization significantly enhanced the quality of learned binary codes.

Conclusions:

• LSDH effectively preserves local data structures, leading to improved hash learning.
• The integration of local semantic awareness and Hamming-isometric quantization offers a significant advancement in deep hashing.
• The proposed method shows strong potential for applications in efficient image retrieval and data storage.