Visual attentional-driven deep learning method for flower recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Forestry Informatization

Background:

• Flower category recognition is a key fine-grained image recognition task.
• Deep Convolutional Neural Networks (DCNNs) show promise but face challenges like limited data, intra-class similarity, and low accuracy in flower recognition.

Purpose of the Study:

• To develop an advanced deep learning model for accurate flower category recognition.
• To address limitations of existing DCNNs in flower recognition, specifically data scarcity and accuracy.

Main Methods:

• Implemented image augmentation (rotation, cropping) to expand the training dataset.
• Proposed a novel attention-driven deep residual neural network incorporating visual attention learning blocks.
• Utilized residual and attention connections to improve network learning and discrimination capabilities.

Main Results:

• The proposed model achieved a recognition accuracy of 85.7% on the public Flowers 17 dataset.
• The attention mechanism enhanced the model's ability to discern subtle differences in flower images.

Conclusions:

• The attention-driven deep learning model effectively improves flower category recognition accuracy.
• This approach offers a promising solution for fine-grained visual classification tasks in computer vision and forestry.