Adversarial Binary Mutual Learning for Semi-Supervised Deep Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Deep hashing methods show promise but struggle with limited labeled data and ineffective distance constraints.
• Existing methods often fail to leverage abundant unlabeled data, hindering performance.
• Traditional pairwise loss functions inadequately enforce desired distance separations for similar/dissimilar pairs.

Purpose of the Study:

• To develop a novel semi-supervised deep hashing model that effectively utilizes both labeled and unlabeled data.
• To introduce a new loss function that explicitly enforces discriminative feature representations.
• To improve the performance of deep hashing methods by addressing limitations in data utilization and feature learning.

Main Methods:

• Proposed adversarial binary mutual learning (ABML) model with a generative (GH) and discriminative (DH) network.
• Employed adversarial learning (AL) for knowledge transfer from unlabeled data to the discriminative model.
• Introduced Weibull cross-entropy loss (WCE) using Weibull distribution for enhanced distance discrimination.

Main Results:

• The ABML model effectively learns from both labeled and unlabeled data through adversarial mutual learning.
• The WCE loss successfully distinguishes subtle distance differences and enforces desired separations.
• Experiments on multiple datasets (CIFAR-10, MNIST, ImageNet-10, NUS-WIDE, ImageNet) show significant performance improvements over state-of-the-art methods.

Conclusions:

• The proposed ABML model with WCE loss overcomes the limitations of existing semi-supervised deep hashing methods.
• The approach achieves more semantic and discriminative features, leading to superior performance in hashing-based retrieval.
• This work offers a significant advancement in efficient and effective large-scale image retrieval using deep hashing.