Fine-Grained Butterfly Recognition via Peer Learning Network with Distribution-Aware Penalty Mechanism.

Area of Science:

• Agricultural Science
• Computer Vision
• Biodiversity Studies

Background:

• Fine-grained species recognition is crucial for intelligent agriculture and biodiversity research.
• Challenges include subtle inter-class differences and long-tailed data distributions.
• Existing methods struggle with imbalanced datasets, leading to biased recognition.

Purpose of the Study:

• To develop a novel method for accurate fine-grained species recognition in challenging, long-tailed datasets.
• To mitigate bias and variance issues inherent in imbalanced sample distributions.
• To enhance recognition accuracy for both common and rare species.

Main Methods:

• A peer learning network utilizing a two-stream ResNeSt-50 backbone is proposed.
• A knowledge exchange strategy selects samples for model updates.
• A distribution-aware penalty mechanism addresses long-tailed data bias.
• A large-scale butterfly dataset (Butterfly-914) with 72,152 images across 914 species was created.

Main Results:

• The proposed method significantly improved recognition accuracy on the Butterfly-914 dataset.
• Achieved a Top-1 accuracy rate of 86.2% on the challenging butterfly dataset.
• Demonstrated effective mitigation of bias and variance in long-tailed distributions.

Conclusions:

• The peer learning network with a distribution-aware penalty mechanism offers a robust solution for fine-grained species recognition.
• The method shows promise for applications in agricultural species identification and insect monitoring.
• The Butterfly-914 dataset provides a valuable resource for advancing research in this domain.