Content-Aware Rectified Activation for Zero-Shot Fine-Grained Image Retrieval.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Fine-grained image retrieval typically learns from seen subcategories, limiting performance in zero-shot settings.
• Salient features learned from seen categories can restrain the discovery of diverse clues needed for unseen categories.

Purpose of the Study:

• To propose a novel Content-aware Rectified Activation model for improved fine-grained image retrieval, especially in zero-shot scenarios.
• To enable models to suppress salient region activation while preserving discrimination and spreading it to non-salient areas.

Main Methods:

• Developed a Content-aware Rectified Activation model that suppresses salient region activation and spreads it to adjacent non-salient regions.
• Introduced a content-aware rectified prototype (CARP) as a channel-wise activation upper bound to obtain rectified features.
• Proposed two regularizations: semantic coherency constraint and feature-navigated constraint to balance feature discrimination and suppression.

Main Results:

• The proposed model effectively mines diverse discriminative features for retrieving unseen subcategories.
• Experimental results on fine-grained and product retrieval benchmarks show consistent outperformance over state-of-the-art methods.
• Demonstrated the model's ability to adaptively balance discrimination and suppression power.

Conclusions:

• The Content-aware Rectified Activation model offers a significant advancement in fine-grained image retrieval, particularly for zero-shot tasks.
• By focusing on diverse features and rectifying salient region activations, the model enhances retrieval accuracy for previously unseen object categories.
• The proposed approach provides a robust framework for future research in discriminative feature learning for image retrieval.