Real-Time Detection Algorithm for Kiwifruit Canker Based on a Lightweight and Efficient Generative Adversarial...

Area of Science:

• Agricultural science
• Computer vision
• Plant pathology

Background:

• Traditional plant disease detection relies on subjective human inspection, lacking accuracy and real-time capabilities.
• Computer vision for disease detection requires extensive specialized data, which is often difficult to obtain due to disease seasonality and rarity.

Purpose of the Study:

• To address data limitations in plant disease detection by developing a high-precision method for generating realistic disease images.
• To improve the accuracy and efficiency of detecting kiwifruit trunk bacterial canker (Pseudomonas syringae pv. actinidiae).

Main Methods:

• Developed a lightweight image generative model incorporating depth-wise separable convolutions and a novel GASLE module for realistic and diverse image generation.
• Utilized an AdaMod optimizer to enhance network convergence.
• Employed the YOLOv8 model for real-time disease detection, evaluating the generative model's effectiveness.

Main Results:

• The generative model achieved a Fréchet Inception Distance (FID) of 84.18, outperforming existing models like FastGAN and ProjectedGAN.
• The YOLOv8 detection model achieved a mean Average Precision (mAP@0.5) of 87.17%, a significant improvement over the original algorithm.
• Real-time detection capability was demonstrated, with the YOLOv8 model processing images in just 0.01 seconds.

Conclusions:

• The proposed generative model effectively creates diverse and realistic plant disease images, overcoming data scarcity issues.
• The integrated approach significantly enhances the accuracy and efficiency of automated plant disease detection systems.
• This strategy offers a robust solution for image generation and disease detection, applicable to various plant diseases.