Bidirectional Mapping Coupled GAN for Generalized Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized Zero-Shot Learning (GZSL) methods depend on synthesized features for recognizing both seen and unseen data.
• Existing GZSL approaches often fail to leverage unseen class semantics, leading to suboptimal performance.
• Preserving the distinction between seen and unseen classes is critical but often overlooked in current GZSL models.

Purpose of the Study:

• To develop a novel GZSL method that effectively utilizes both seen and unseen class semantics for improved feature synthesis.
• To address the limitations of existing methods by learning a joint distribution that preserves seen-unseen class distinction.
• To enhance the accuracy and reduce bias towards seen classes in GZSL.

Main Methods:

• Proposed a bidirectional mapping coupled generative adversarial network (BMCoGAN) model.
• Integrated Wasserstein generative adversarial optimization for supervised joint distribution learning.
• Designed a specialized loss function to retain distinctive seen-unseen class information and mitigate bias towards seen classes.

Main Results:

• BMCoGAN effectively utilizes unseen class semantics alongside seen class semantics through strong visual-semantic coupling.
• The proposed loss optimization successfully reduces bias towards seen classes and enhances the distinctiveness of synthesized features.
• Evaluations on benchmark datasets demonstrate superior performance of BMCoGAN compared to contemporary GZSL methods.

Conclusions:

• The proposed BMCoGAN model offers a significant advancement in GZSL by effectively learning joint distributions and preserving class distinctiveness.
• Leveraging unseen class semantics and employing Wasserstein optimization are key to achieving superior GZSL performance.
• The method shows strong potential for real-world applications requiring recognition of novel object categories.