Feature Enhancement Module Based on Class-Centric Loss for Fine-Grained Visual Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained visual classification (FGVC) is challenging due to subtle inter-class variations.
• Existing methods often struggle to capture discriminative features necessary for distinguishing similar categories.

Purpose of the Study:

• To introduce a novel, versatile feature enhancement module (FEM) for FGVC tasks.
• To improve the representation of fine-grained features through pixel-level enhancement and weighted fusion.
• To develop a class-centric loss function for better sample-class alignment.

Main Methods:

• The proposed plug-and-play module integrates with various backbone architectures (CNNs and Transformers).
• It outputs pixel-level feature maps and applies weighted fusion of filtered features.
• A class-centric loss function is introduced, aligning samples with target class centers and repelling them from similar non-target classes.
• Soft labels are utilized to prevent overfitting and enhance generalization.

Main Results:

• The approach consistently improved accuracy across diverse mainstream backbone architectures.
• Significant performance gains were observed in fine-grained feature representation.
• The method achieved state-of-the-art accuracy on the NABirds (NAB) dataset.
• Top accuracy was also obtained on a proprietary lock cylinder dataset.

Conclusions:

• The feature enhancement module offers a versatile and robust solution for FGVC.
• The class-centric loss function effectively optimizes feature discrimination.
• The proposed method demonstrates broad applicability and superior performance on challenging fine-grained datasets.