Unsupervised t-Distributed Video Hashing and Its Deep Hashing Extension

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Large-scale video retrieval demands efficient indexing and search mechanisms.
• Unsupervised learning methods are crucial for handling vast, unlabeled video datasets.
• Existing hashing techniques often struggle with robustness and scale variations.

Purpose of the Study:

• To propose novel unsupervised hashing algorithms (t-USMVH and t-UDH) for large-scale video-to-video retrieval.
• To enhance the robustness of unsupervised learning by integrating multiple feature representations.
• To improve the accuracy of hash code generation by reducing sensitivity to scale changes.

Main Methods:

• Developed t-USMVH, combining diverse feature representations using continuous relevance and discrete neighbor scores.
• Employed Student t-distribution for similarity estimation between keyframe hash codes, mitigating scale sensitivity.
• Extended t-USMVH to unsupervised deep hashing (t-UDH) using deep neural networks for robust training.

Main Results:

• The proposed t-USMVH and t-UDH methods demonstrate improved accuracy in preserving unsupervised similarity structures.
• Evaluations on public video collections show superior performance compared to classical and state-of-the-art methods.
• The algorithms exhibit efficiency and effectiveness for large-scale video retrieval tasks.

Conclusions:

• t-USMVH and t-UDH offer a robust and effective solution for unsupervised large-scale video retrieval.
• The integration of Student t-distribution and multi-feature fusion significantly enhances hashing performance.
• These methods advance the field of unsupervised deep hashing for complex multimedia data.