Scalable Deep Hashing for Large-scale Social Image Retrieval

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hashing is crucial for efficient large-scale image retrieval due to low storage and high computation speed.
• Supervised deep hashing methods achieve high performance but are limited by scalability due to reliance on manual labels.
• Unsupervised deep hashing offers scalability but suffers from limited performance due to lack of semantic guidance and relaxed optimization.

Purpose of the Study:

• To propose a Scalable Deep Hashing (SCADH) framework for enhanced hash code generation in social image retrieval.
• To leverage weak, free supervision from user tags commonly associated with social images.
• To improve retrieval performance by enhancing the discriminative capability of image hash codes.

Main Methods:

• Developed a unified scalable deep hash learning framework integrating image representation and hash function learning.
• Utilized discriminative user tags as weak supervision for semantic guidance.
• Implemented a discrete hash optimization method using Augmented Lagrangian Multiplier to avoid quantization loss.

Main Results:

• The proposed SCADH framework demonstrates superior performance compared to existing shallow and deep hashing techniques.
• Experiments on two standard social image datasets validate the effectiveness of the SCADH approach.
• The method successfully enhances the discriminative capability of image hash codes using social tags.

Conclusions:

• SCADH offers a scalable and effective solution for social image retrieval by utilizing readily available user tags.
• The discrete hash optimization method preserves binary quantization information, leading to improved hash code quality.
• This approach overcomes the limitations of both supervised and unsupervised deep hashing methods.