BinoHeM: Binocular Singular Hellinger Metametric for Fine-Grained Few-Shot Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Meta-metric learning excels in coarse-grained few-shot tasks but struggles with fine-grained classification.
• Fine-Grained Few-Shot Classification (FGFSC) requires subtle feature extraction, a challenge for current models.
• Human visual systems demonstrate robust meta-learning for fine-grained recognition.

Purpose of the Study:

• To introduce a novel human-like meta-metric learning paradigm inspired by binocular vision.
• To address limitations of existing metametrics in fine-grained few-shot scenarios.
• To develop advanced variants for improved performance in FGFSC.

Main Methods:

• Pioneered the Binocular Singular Hellinger Metametric (BinoHeM) paradigm.
• Incorporated symmetric binocular feature encoding and recognition mechanisms.
• Introduced BinoHeM-KDL and BinoHeM-MTL variants using knowledge distillation and meta-transfer learning.

Main Results:

• Demonstrated high accuracy and robust generalization on four FGFSC benchmarks.
• Achieved superior performance compared to state-of-the-art algorithms.
• Validated through extensive comparative and ablation experiments.

Conclusions:

• The BinoHeM paradigm offers a novel and effective approach to fine-grained few-shot classification.
• Human-inspired meta-metric learning shows significant potential for complex visual recognition tasks.
• The proposed methods, BinoHeM-KDL and BinoHeM-MTL, represent advancements in FGFSC.