Learnable Central Similarity Quantization for Efficient Image and Video Retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Traditional data-dependent hashing methods capture local data distributions, leading to inefficiencies and low collision rates.
• Existing methods struggle with capturing global data relationships effectively.

Purpose of the Study:

• To introduce a novel global similarity metric, central similarity, for enhanced hash learning.
• To develop efficient methods for generating well-separated hash centers.
• To propose and evaluate central similarity quantization (CSQ) and its variant (CSQLC) for deep hash function generation.

Main Methods:

• Central similarity metric: encourages similar data hash codes to approach common centers and dissimilar ones to diverge.
• Hash center generation: utilizes Hadamard matrices and Bernoulli distributions (data-independent) or learns from data representations (data-dependent).
• Central Similarity Quantization (CSQ): optimizes central similarity with respect to hash centers for high-quality deep hash functions.

Main Results:

• CSQ and CSQLC demonstrate significant improvements in retrieval performance, with mean average precision (mAP) gains of 3%-20% over state-of-the-art methods.
• The methods effectively generate cohesive hash codes for similar data pairs and dispersed codes for dissimilar pairs.
• Experiments on large-scale image and video retrieval tasks validate the proposed approach's effectiveness.

Conclusions:

• Central similarity offers a superior global metric for hash learning, overcoming limitations of local distribution-based methods.
• The proposed CSQ and CSQLC methods provide efficient and effective solutions for deep hash function learning in multimedia retrieval.
• The approach significantly enhances retrieval accuracy and efficiency in image and video hashing applications.