Exemplar-Based, Semantic Guided Zero-Shot Visual Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot recognition is a challenging area due to the absence of direct supervision.
• Existing methods often overlook individual image distinctiveness, focusing on class-level modeling.
• Semantic information is commonly used as a bridge in zero-shot recognition tasks.

Purpose of the Study:

• To propose a novel exemplar-based, semantic guided zero-shot recognition (EBSG) method.
• To leverage both visual and semantic information at the individual image level.
• To enhance zero-shot recognition by considering distinctive image characteristics.

Main Methods:

• Developed an exemplar-based, semantic guided zero-shot recognition (EBSG) method.
• Trained a visual sub-model to differentiate individual images across classes.
• Trained a semantic sub-model to distinguish images based on unique semantics.
• Concatenated outputs from visual and semantic sub-models for image representation.
• Learned an image classification model using visual similarity and semantic consistency.

Main Results:

• The proposed EBSG method demonstrated effectiveness in zero-shot recognition.
• Experiments were conducted on four widely used benchmark datasets.
• The EBSG method showed superior performance compared to existing approaches.

Conclusions:

• The EBSG method effectively utilizes individual image visual and semantic information.
• Considering distinctive image characteristics is crucial for improving zero-shot recognition.
• The proposed approach offers a promising direction for future zero-shot recognition research.