Fine-Grained Recognition With Learnable Semantic Data Augmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Fine-grained image recognition is challenging due to subtle visual differences within meta-categories.
• Standard image augmentation can destroy critical discriminative features in fine-grained tasks.

Purpose of the Study:

• To address the loss of discriminative visual cues in fine-grained image recognition.
• To propose a feature-level data augmentation technique that preserves subtle details.

Main Methods:

• Developed a feature-level data diversification strategy by translating image features along semantically meaningful directions.
• Introduced a covariance prediction network to estimate semantic directions and adapt to intra-class variations.
• Jointly optimized the covariance prediction and classification networks using meta-learning.

Main Results:

• Significantly improved generalization performance across multiple popular classification networks (ResNets, DenseNets, EfficientNets, RegNets, ViT).
• Achieved state-of-the-art results on the CUB-200-2011 benchmark when combined with existing methods.
• Demonstrated effectiveness on four competitive fine-grained recognition datasets.

Conclusions:

• Feature-level semantic data augmentation is effective for fine-grained image recognition.
• The proposed covariance prediction network successfully captures intra-class variations.
• The method offers a robust solution for enhancing discriminative feature learning.