Plant Disease Detection Using Generated Leaves Based on DoubleGAN.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Plant disease detection is crucial for agriculture.
• Unbalanced datasets of unhealthy plant leaf images hinder accurate disease detection.
• Existing methods struggle with limited and imbalanced data.

Purpose of the Study:

• To develop a method for generating high-resolution, realistic images of unhealthy plant leaves.
• To balance imbalanced datasets for improved plant disease detection.
• To enhance the accuracy of plant species and disease recognition systems.

Main Methods:

• Proposed DoubleGAN (a double generative adversarial network) comprising two stages.
• Stage 1: Utilized Wasserstein generative adversarial network (WGAN) for pretraining and generating 64x64 pixel unhealthy leaf images.
• Stage 2: Employed super-resolution generative adversarial network (SRGAN) to upscale images to 256x256 pixels, expanding the dataset.

Main Results:

• DoubleGAN generated clearer images compared to Deep Convolutional Generative Adversarial Network (DCGAN).
• The dataset expanded using DoubleGAN led to improved recognition accuracy.
• Achieved 99.80% accuracy for plant species recognition and 99.53% for disease recognition.

Conclusions:

• DoubleGAN effectively generates high-resolution unhealthy plant leaf images from limited samples.
• The proposed method successfully balances imbalanced datasets.
• The enhanced dataset significantly improves the performance of plant disease and species recognition models.