Parameter-Efficient Tuning for Fine-Grained Recognition via Channel-Wise Importance Equalization and Diversity...

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Parameter-efficient tuning (PET) shows promise for vision tasks but struggles with fine-grained recognition due to channel saliency and feature redundancy.
• Existing PET methods often overlook the over-concentration of channel-wise saliency and feature redundancy in pre-trained models.

Purpose of the Study:

• To introduce a novel parameter-efficient tuning approach, FG-PET, specifically designed to enhance fine-grained recognition.
• To address the limitations of existing PET methods in handling channel saliency and feature redundancy for fine-grained tasks.

Main Methods:

• FG-PET utilizes a Channel-wise Importance Equalization (CIE) module to mitigate over-concentrated saliency by balancing channel importance.
• An Efficient Navigator for Diversity (EFIND) module is employed, incorporating center-based loss and orthogonal constraints to reduce feature redundancy and encourage diverse feature exploration.

Main Results:

• FG-PET significantly improves performance on five public fine-grained recognition benchmarks using various Vision Transformer (ViT) models.
• The proposed method demonstrates strong generalization capabilities, also enhancing performance on general classification tasks.

Conclusions:

• FG-PET effectively addresses key limitations in existing parameter-efficient tuning methods for fine-grained recognition.
• The approach enhances the ability of models to capture subtle visual differences crucial for fine-grained tasks and shows broad applicability.