TelecomNet: Tag-Based Weakly-Supervised Modally Cooperative Hashing Network for Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• User-tagged images offer abundant data and rich semantic information for image retrieval.
• Existing tagging data presents challenges like noise, vagueness, and incompleteness.
• Current deep hashing methods often struggle with imperfect tagging data.

Purpose of the Study:

• To develop a novel weakly-supervised deep hashing framework to effectively utilize noisy user-tagged images for image retrieval.
• To propose two specific formulations, TelecomNet and GTelecomNet, within this framework.
• To enhance the characterization of similarity relationships between images using semantic information from tags.

Main Methods:

• A two-stage framework: weakly-supervised pre-training followed by supervised fine-tuning.
• TelecomNet learns an observed semantic embedding from tags to guide hashing.
• GTelecomNet utilizes a novel semantic network for more precise semantic information extraction.
• Careful optimization problem design to leverage both tagging information and image content.

Main Results:

• Empirical results on real-world datasets demonstrate significant performance improvements.
• The proposed methods effectively address noise, vagueness, and incompleteness in tagging data.
• The framework enhances the performance of state-of-the-art deep hashing methods.

Conclusions:

• The weakly-supervised deep hashing framework offers a robust solution for leveraging imperfect user-tagged images.
• TelecomNet and GTelecomNet provide effective strategies for incorporating semantic information into hashing learning.
• This approach significantly advances the capabilities of deep hashing for image retrieval.