CNN and ResNet50 Performance Comparison for Maize Leaf Disease Detection

Abstract

Supervision of crop health and increase of agricultural output depend on early disease detection and precise classification of illnesses in maize (corn) leaves. This work represents a thorough comparative analysis of two models, Convolutional Neural Networks (CNN) and ResNet50, for maize leaf disease identification. A CNN model including several convolutional and pooling layers and ResNet50, a pretrained deep learning model well-known for its residual learning architecture were used for classification. With an $\mathbf{9 1 \%}$ accuracy, the CNN model proved efficient and with less computing demand. ResNet50, on the other hand, benefited from its deeper architecture and usage of residual connections that help to solve the vanishing gradient issue, therefore attaining a better accuracy of 92%. Data augmentation techniques were used to improve generalisation of models. To evaluate the models, accuracy, precision, recall, and F1-score were calculated among other performance measures. Though it needed more computing power and longer training cycles, ResNet50 exceeded the CNN model in terms of accuracy and robustness. The paper emphasises the trade-offs between model complexity and performance, implying that although ResNet50 provides better accuracy, in contexts with limited resources the CNN model might be used. This comparison offers insightful analysis of CNN and ResNet50 model suitability for practical agricultural diagnostics. Future research will concentrate on improving these models and including them into mobile applications to help farmers in prompt and efficient disease control, therefore supporting sustainable agricultural methods.

Topics & Concepts