Plant Disease Recognition from Leaf Images Using Convolutional Neural Network

Abstract

Automated plant disease recognition from leaf images of a plant is important in the field of agriculture. To achieve this, a convolutional neural network (ConvNet)-based classifier is proposed in this paper. The existing methods used the general CNN architectures such as AlexNet and VGGNet for the disease recognition by retraining it whereas the proposed CNN architecture contains a few convolution layers against the existing CNN models. The number of learnable parameters in the existing networks are 30 times higher compared to the proposed model. The proposed ConvNet is trained on a subset of the publically available dataset that contains 24,249 images of diseased and healthy plant leaves of the crops cultivated in India. The image samples in the dataset are modified by thresholding the original images to remove the background. Rectified linear unit (ReLu) activation function is used across all layers. The sparse categorical cross entropy loss function is considered with Adam optimization algorithm to fine tune the network parameters by minimizing the loss function. The proposed model achieved a training accuracy of 96.39%, validation accuracy of 90.97%, and testing accuracy of 90.59%. The performance of the proposed model is comparable to the existing methods, however, with reduced number of parameters. Due to the reduction in model parameters to a larger extent, the proposed model could be deployed in a resource constrained edge computing devices for a real-time processing.