Hybrid Granularities Transformer for Fine-Grained Image Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained image recognition requires identifying subtle features, a challenge for models focusing on prominent elements.
• Intra-class variation and inter-class similarity in fine-grained datasets complicate discriminative feature extraction.

Purpose of the Study:

• To introduce two novel, lightweight modules designed to enhance feature extraction in fine-grained image classification.
• To enable networks to discover and utilize subtle, detailed information within images for improved classification accuracy.

Main Methods:

• Patches Hidden Integrator (PHI): Randomly replaces image patches with those from the same class to encourage diverse feature learning and prevent over-reliance on single features.
• Consistency Feature Learning (CFL): Aggregates patch tokens, fuses local features with class tokens, and uses inconsistency loss to focus on salient regions.

Main Results:

• Achieved high accuracy on benchmark datasets: 91.6% on CUB-200-2011, 92.7% on Stanford Dogs, and 99.5% on Oxford 102 Flowers.
• Demonstrated competitive performance compared to existing state-of-the-art methods in fine-grained image recognition.

Conclusions:

• The proposed PHI and CFL modules effectively address the challenges of fine-grained image recognition by capturing detailed and discriminative features.
• These lightweight modules offer a promising approach for enhancing deep learning models in specialized image classification tasks without increasing training time.