Visual feature-based multi-scale hybrid attention network for fine-grained Hawthorn varieties identification.

Area of Science:

• Botany
• Computer Science
• Agricultural Science

Background:

• Hawthorn (Crataegus spp.) is valued for cardiovascular benefits, but variety identification is difficult due to cultivation variations.
• Accurate species authentication is essential for quality control and efficacy of Hawthorn products.
• Current identification methods face challenges in distinguishing between numerous Hawthorn varieties.

Purpose of the Study:

• To develop an automated and accurate method for Hawthorn variety identification using visual features.
• To enhance species authentication for the economically important Hawthorn crop.
• To improve the precision of classifying fine-grained botanical images.

Main Methods:

• A multi-scale hybrid deep learning model was proposed to integrate local and global image features.
• A novel spatial local attention mechanism was incorporated to improve recognition of fine-grained details.
• Customized loss functions were designed to reduce low-frequency features in images, enhancing classification.

Main Results:

• The proposed deep learning model achieved superior performance in Hawthorn identification compared to existing state-of-the-art methods.
• Experiments on dedicated and public datasets validated the model's accuracy and robustness.
• The model effectively captured both local details and global context in Hawthorn images.

Conclusions:

• The developed visual feature-based deep learning method offers a reliable solution for Hawthorn species authentication.
• This approach significantly advances the capability for precise identification of Hawthorn varieties.
• The findings have implications for quality control, research, and commercial applications of Hawthorn.