Convolutional Fine-Grained Classification With Self-Supervised Target Relation Regularization.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Traditional target coding schemes in deep representation learning lack flexibility for inter-class correlations.
• Existing methods are sensitive to sparse and imbalanced data distributions, hindering fine-grained visual classification.

Purpose of the Study:

• To introduce a novel target coding scheme, dynamic target relation graphs (DTRG), for improved fine-grained visual classification.
• To enhance feature regularization by generating structural outputs from input images.

Main Methods:

• Developed dynamic target relation graphs (DTRG) as an auxiliary feature regularization technique.
• Implemented online computation of class-level feature centers to generate cross-category distances in representation space.
• Introduced randomness into soft construction of DTRG, inspired by mixup data augmentation, to explore relation diversity.

Main Results:

• Demonstrated the effectiveness of DTRG on diverse visual classification benchmarks.
• Achieved state-of-the-art performance on three popular fine-grained object recognition datasets.
• Showcased superior robustness against sparse and imbalanced data distributions.

Conclusions:

• Dynamic target relation graphs (DTRG) offer a flexible and robust approach to fine-grained visual classification.
• The proposed method effectively addresses limitations of traditional target coding schemes, particularly with challenging data distributions.
• DTRG facilitates learning discriminative features by minimizing intra-class variations and exploiting inter-class dependencies.