Pepper leaf disease recognition based on enhanced lightweight convolutional neural networks.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Convolutional neural networks (CNNs) are crucial for identifying pepper leaf diseases.
• Existing CNN models face challenges with accuracy and computational demands on portable devices.
• Efficient leaf disease recognition is vital for large-scale pepper farming.

Purpose of the Study:

• To develop an enhanced, lightweight CNN model for accurate and efficient pepper leaf disease identification.
• To address the limitations of existing models in terms of computational performance and memory usage.
• To enable the deployment of CNNs on embedded systems for real-time agricultural monitoring.

Main Methods:

• An enhanced lightweight model based on the GoogLeNet architecture was developed.
• The Inception structure was compressed to reduce parameters and increase recognition speed.
• Spatial pyramid pooling was integrated to combine local and global features.
• The model was trained on a dataset of 9183 images across 6 pepper disease types.

Main Results:

• The enhanced model achieved 97.87% accuracy, outperforming GoogLeNet variants by 6%.
• Model memory requirement was reduced to 10.3 MB (52.31%-86.69% reduction).
• Average inference time decreased significantly compared to AlexNet, ResNet-50, and MobileNet-V2.

Conclusions:

• The proposed enhanced model offers superior accuracy and computational efficiency for pepper leaf disease identification.
• Its lightweight design makes it suitable for deployment on embedded portable devices.
• This advancement holds potential for improving productivity and disease management in the pepper farming industry.