[Paper Review] Semi-supervised Hashing for Semi-Paired Cross-View Retrieval

Jun Yu, Xiao‐Jun Wu+1 more|arXiv (Cornell University)|Jun 19, 2018

Advanced Image and Video Retrieval Techniques10 references3 citations

TL;DR

This paper proposes a semi-supervised hashing framework for semi-paired cross-view retrieval, jointly learning feature extraction and classification using partial label information. By leveraging both labeled and unlabeled data, the method achieves superior retrieval accuracy on two benchmark datasets compared to state-of-the-art approaches.

ABSTRACT

Recently, hashing techniques have gained importance in large-scale retrieval tasks because of their retrieval speed. Most of the existing cross-view frameworks assume that data are well paired. However, the fully-paired multiview situation is not universal in real applications. The aim of the method proposed in this paper is to learn the hashing function for semi-paired cross-view retrieval tasks. To utilize the label information of partial data, we propose a semi-supervised hashing learning framework which jointly performs feature extraction and classifier learning. The experimental results on two datasets show that our method outperforms several state-of-the-art methods in terms of retrieval accuracy.

Motivation & Objective

• Address the limitation of existing cross-view hashing methods that assume fully paired data, which is uncommon in real-world applications.
• Enable effective retrieval in scenarios where only partial data pairs are available.
• Leverage limited labeled data and abundant unlabeled data to improve hashing function learning.
• Develop a unified framework that jointly optimizes feature extraction and classifier learning for improved retrieval performance.
• Achieve higher retrieval accuracy than state-of-the-art methods on semi-paired cross-view retrieval tasks.

Proposed method

• Propose a semi-supervised hashing framework that integrates feature learning and classifier learning in a joint optimization process.
• Utilize partial label information from a subset of data to guide the learning of hashing functions.
• Formulate the learning objective to preserve semantic similarity between paired samples while aligning features across views.
• Introduce a classifier head that is trained alongside the hashing network to enhance discriminative capacity.
• Optimize the model end-to-end using a loss function combining contrastive, classification, and regularization terms.
• Apply deep neural networks to extract deep features from multiple views, followed by binary code generation.

Experimental results

Research questions

• RQ1 Can a semi-supervised hashing framework effectively handle semi-paired cross-view retrieval where only partial data pairs exist?
• RQ2 How does joint learning of features and classifiers improve retrieval accuracy compared to isolated learning?
• RQ3 To what extent does using partial label information enhance performance in low-supervision settings?
• RQ4 Does the proposed method outperform fully paired or unsupervised baselines in semi-paired scenarios?
• RQ5 How robust is the method across different datasets with varying levels of data pairing?

Key findings

• The proposed method achieves higher mean average precision (mAP) than state-of-the-art methods on both benchmark datasets.
• The model demonstrates significant performance gains when only a small fraction of data are labeled, proving the effectiveness of semi-supervised learning.
• Joint optimization of feature learning and classification leads to more discriminative hash codes compared to separate training.
• The method maintains strong performance even with limited labeled data, showing robustness to label scarcity.
• Experimental results confirm that the framework outperforms existing cross-view hashing methods in semi-paired retrieval settings.
• The ablation study verifies that both the semi-supervised loss and joint learning components contribute significantly to performance.