A Comparative Study of Different Architectural Models of CNN for Plant Leaf Disease Detection

Abstract

Purpose – From the last few decades, pattern recognition has become an emerging task of machine learning and image processing with robust integration. This paper provides a comparative study of different plant leaf disease detection techniques of the CNN model in the domain of image processing.

Method – In this paper, we compared three architectural models of CNN namely, AlexNet, VGG16Net, and ResNet for plant disease detection. AlexNet has five convolution layers followed by three fully connected layers. VGG uses a small receptive field followed by a ReLu unit and it has three fully connected layers. ResNet works on skip connection and it passes input data through the weight layer processing by model function.

Results – ResNet provides an effective result with 100 epoch iterations of dataset training and validation. ResNet achieved higher training and validation accuracy than AlexNet and VGG16Net models. ResNet has also achieved less training and validation loss. Finally, the experimental results have shown that ResNet is better than AlexNet and VGG16Net models.

Conclusion – In this study, we concluded that the residual network i.e. ResNet is showing better results than AlexNet and VGG16Net. Finally, the comparative experimental results have shown that ResNet provides effective output with 100 epochs.

Recommendations – The recognition rate of ResNet needs to be tested by increasing the number of epoch’s iterations and adding more and new leaf data for training and testing datasets for future work. In future research, we recommended the development of an Android-based mobile App for plant leaf disease detection useful for farmers.

Research Implications – Farmers can easily operate this system on their smartphones with a few days technical training given by expert professionals to detect plant leaf disease.