Transductive Zero-Shot Hashing for Multilabel Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hash coding is crucial for approximate nearest neighbor search in large-scale image retrieval.
• Existing zero-shot hashing (ZSH) methods primarily address single-label images, lacking capabilities for multilabel scenarios.
• Retrieving unseen images with undefined semantic labels presents a significant challenge in image retrieval.

Purpose of the Study:

• To propose a novel transductive zero-shot hashing (ZSH) method for multilabel unseen image retrieval.
• To develop a technique capable of predicting labels for unseen images in a multilabel context.
• To enhance the performance of image retrieval systems dealing with complex, multilabel datasets.

Main Methods:

• A visual-semantic bridge is established using instance-concept coherence ranking on seen data.
• A hashing model is trained using both seen and unseen data with pairwise similarity loss and focal quantization loss.
• The proposed method is a transductive approach, leveraging both labeled and unlabeled data during training.

Main Results:

• The novel ZSH method demonstrates superior performance in multilabel unseen image retrieval compared to existing methods.
• Extensive evaluations on three popular multilabel datasets confirm the effectiveness of the proposed approach.
• The method successfully bridges the gap between visual features and semantic labels for unseen images.

Conclusions:

• The developed transductive ZSH method is effective for multilabel unseen image retrieval.
• The proposed approach significantly improves retrieval accuracy for complex image datasets.
• This work opens new avenues for zero-shot learning in large-scale image retrieval systems.