Fine-Grained Image Recognition with Bio-Inspired Gradient-Aware Attention.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biologically Inspired Computing

Background:

• Fine-grained image recognition is challenging due to subtle inter-class and intra-class variations.
• Traditional methods struggle with background noise and feature degradation.
• Human visual systems effectively focus on discriminative regions.

Purpose of the Study:

• To develop a novel attention mechanism for improved fine-grained image recognition.
• To address limitations of conventional approaches in handling background interference and feature degradation.
• To mimic human visual system's focus on discriminative regions.

Main Methods:

• Proposed a bio-inspired gradient-aware attention mechanism.
• Modeled gradient information to guide attention, mimicking biological edge sensitivity.
• Enhanced discrimination between global structures and local details.

Main Results:

• Achieved Top-1 accuracy of 92.9% on CUB-200-2011.
• Achieved Top-1 accuracy of 90.5% on iNaturalist2018.
• Achieved Top-1 accuracy of 93.1% on nabbirds.
• Achieved Top-1 accuracy of 95.1% on Stanford Cars.

Conclusions:

• The proposed gradient-aware attention mechanism significantly improves fine-grained image recognition.
• The bio-inspired approach effectively enhances feature discrimination by leveraging gradient information.
• The method demonstrates superior performance across multiple benchmark datasets.