Unsupervised Deep Video Hashing via Balanced Code for Large-Scale Video Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Large-scale video similarity search is crucial for efficient information retrieval.
• Existing video hashing methods often require supervision or lack effectiveness in generating compact binary codes.

Purpose of the Study:

• To propose an unsupervised deep hashing framework (UDVH) for large-scale video similarity search.
• To learn compact and effective binary codes for videos without relying on labeled data.

Main Methods:

• Developed Unsupervised Deep Video Hashing (UDVH), a framework integrating discriminative video representation with optimal code learning.
• Employed an efficient alternating approach for objective function optimization.
• Incorporated feature clustering and a novel binarization method preserving neighborhood structure.
• Introduced a specific rotation technique to balance feature dimension variance for improved quantization.

Main Results:

• UDVH demonstrates superior performance compared to state-of-the-art methods across various evaluation metrics.
• The framework successfully generates compact yet effective binary codes for video representation.
• Experimental results on three benchmark datasets validate the effectiveness of UDVH.

Conclusions:

• UDVH is a highly effective unsupervised method for large-scale video similarity search.
• The proposed framework offers practical solutions for real-world video retrieval applications.
• UDVH's novel approach to feature representation and binarization significantly advances the field of video hashing.