Corn Plant Disease Detection Using Deep Learning

Abstract

Detection of diseases on corn leaves based on images requires a Convolutional Neural Network (CNN) model capable of accurately recognizing visual patterns because the symptoms often appear similar across disease classes. A CNN with a transfer learning approach was used due to its ability to automatically extract visual features. In its implementation, we compared three CNN architectures namely VGG16, INCEPTION-V3, and DenseNet to identify the most effective architecture. This comparison is necessary because each model differs in layer depth, feature extraction strategy, and complexity, which can influence model performance on the corn leaf dataset. The training process utilized the Adam optimizer with a learning rate of 0.0001. The results indicate that VGG16 achieved a training accuracy of 90% and validation accuracy of 91%, INCEPTION-V3 achieved a training accuracy of 92.10% and validation accuracy of 93.02%, while DenseNet delivered the highest performance with a training accuracy of 95.41% and validation accuracy of 97.05%. Therefore, DenseNet is considered the most effective and has the potential to serve as the basis for developing an image based automatic detection system for corn leaf diseases.