Plant leaf disease recognition based on improved SinGAN and improved ResNet34

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Precision agriculture relies on accurate plant leaf disease identification for crop yield and quality.
• Current deep learning models face challenges with limited datasets and suboptimal accuracy.
• Existing methods struggle with parameter efficiency and feature extraction.

Purpose of the Study:

• To develop an enhanced plant leaf disease identification method using improved generative adversarial networks and convolutional neural networks.
• To address limitations in agricultural datasets and improve the accuracy and efficiency of deep learning models.
• To enhance image quality for better disease detection and diagnosis.

Main Methods:

• Proposed Reconstruction-Based Single Image Generation Network (ReSinGN) using an autoencoder and Convolutional Block Attention Module (CBAM) for image enhancement.
• Incorporated random pixel shuffling in ReSinGN to improve data representation and image quality.
• Developed an improved ResNet34 by integrating CBAM modules, using LeakyReLU activation, and employing transfer learning for faster training.

Main Results:

• ReSinGN demonstrated a 44.6x faster training speed and produced clearer images with a 30.2 improvement in Tenengrad score compared to SinGAN.
• The ReSinGN model achieved an optimal balance between image clarity and distortion.
• The improved ResNet34 achieved high performance metrics for tomato leaf disease identification, including 98.57% accuracy, 96.57% precision, and 98.17% F1 score, surpassing the original ResNet34.

Conclusions:

• The proposed ReSinGN and improved ResNet34 effectively enhance plant leaf disease identification.
• The methods provide a viable solution to challenges posed by limited agricultural datasets and model inaccuracies.
• This work contributes to the advancement of precision agriculture through improved disease detection technologies.