Discriminative Codebook Hashing for Supervised Video Retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Supervised video hashing is crucial for efficient large-scale video retrieval.
• Existing methods often rely on pairwise or triple similarities, focusing on local information, leading to suboptimal retrieval accuracy.
• There is a need for advanced hashing techniques that capture global semantic information for better video retrieval.

Purpose of the Study:

• To propose a novel supervised hashing framework, Discriminative Codebook Hashing (DCH), for large-scale video retrieval.
• To enhance retrieval accuracy by encouraging intra-class compactness and inter-class separability in the Hamming space.
• To address the limitations of existing methods that focus on local information.

Main Methods:

• Developed a discriminative codebook approach with predefined inter-code word distances and Bernoulli distributions for hash bits.
• Employed composite Kullback-Leibler (KL) divergence to align neighborhood structures between high-dimensional and Hamming spaces.
• Optimized the DCH framework using the gradient descent algorithm.

Main Results:

• The proposed DCH framework demonstrated superior performance compared to several state-of-the-art methods.
• Experimental results on three widely used video datasets validated the effectiveness of DCH.
• DCH successfully encouraged samples within the same category to converge to the same code word while maximizing distances between different categories.

Conclusions:

• DCH offers a significant improvement over existing supervised video hashing techniques.
• The method effectively balances local and global information for enhanced video retrieval.
• DCH provides a robust and accurate solution for large-scale supervised video retrieval tasks.