Leaf Recognition Based on Joint Learning Multiloss of Multimodel Convolutional Neural Networks: A Testing for...

Area of Science:

• Computer Vision
• Machine Learning
• Botanical Image Analysis

Background:

• Leaf image recognition is crucial for plant identification and biodiversity monitoring.
• Existing deep convolutional neural network (CNN) models often face limitations in accuracy and generalization.
• Ensemble methods can improve performance but may suffer from redundancy and high computational costs.

Purpose of the Study:

• To develop a novel multimodel CNN (MMCNN) approach for enhanced leaf image recognition.
• To improve the generalization capability and accuracy of leaf recognition systems.
• To reduce the computational cost associated with ensemble classifiers.

Main Methods:

• Image segmentation using the U-Net model to isolate leaf images from backgrounds.
• A multimodel approach combining loss functions from EfficientNet and MobileNet architectures.
• Implementing a joint learning strategy where multiple deep networks cooperate and share knowledge.
• Utilizing a multiloss trade-off strategy to mitigate redundancy in ensemble classifiers.

Main Results:

• The proposed MMCNN approach significantly enhances leaf recognition performance.
• The method outperforms current standard single deep learning networks on various datasets.
• The approach demonstrates improved generalization capability and richer information learning.
• Reduced computational cost compared to traditional ensemble methods was observed.

Conclusions:

• The MMCNN approach offers a more effective and efficient solution for leaf image recognition.
• Combining diverse deep learning models with a joint learning strategy boosts performance.
• This technique provides a robust framework for botanical image analysis and plant identification.