An Improved Deep Residual Convolutional Neural Network for Plant Leaf Disease Detection

Area of Science:

• Computer Science
• Plant Pathology
• Artificial Intelligence

Background:

• Accurate plant disease detection is crucial for food security and agricultural productivity.
• Existing methods often struggle with the complexity and variability of plant diseases.
• Deep learning offers potential for automated and precise disease diagnosis.

Purpose of the Study:

• To develop and evaluate a novel deep residual convolutional neural network (ResNet197) for plant leaf disease detection.
• To assess the performance of ResNet197 against existing architectures and transfer learning techniques.
• To leverage a large, augmented dataset for robust model training.

Main Methods:

• A 197-layer deep residual convolutional neural network (ResNet197) was designed, comprising six layer blocks.
• Extensive data augmentation techniques (scaling, cropping, rotation, etc.) were applied to a dataset of 154,500 images across 103 classes.
• An evolutionary search technique optimized ResNet197's layers and hyperparameters, followed by GPU-accelerated training for 1000 epochs.

Main Results:

• ResNet197 achieved an average classification accuracy of 99.58% on the test dataset.
• The model demonstrated superior performance compared to established ResNet architectures.
• Experimental results indicated significant improvements over recent transfer learning techniques.

Conclusions:

• The proposed ResNet197 model is highly effective for detecting various plant leaf diseases.
• The study highlights the potential of deep residual networks in agricultural applications.
• ResNet197 offers a promising solution for automated plant disease diagnosis.