An Efficient Convolutional Neural Network Based Deep Learning Framework for Rice Leaf Disease Classification

Abstract

Rice is one of the most popular and extensively grown crops. More than fifty percent of the global population consumes it as a staple food. However, a number of diseases affect the quality of crop which also affect the total production of the yield. Identification of such disorders is important for crop management. Traditional disease detection methods are time consuming and susceptible to human error because they rely on personal evaluation by professionals. In our work, we have built a task specific system. This system uses a convolutional neural network based framework for detection and classification of rice leaf diseases by looking at leaf pictures. Two datasets are used in this work. The rice leaf disease dataset contains 16000 images which are distributed into four classes with 4000 images in each class and additionally 139 images of brown spot class from A dataset of rice leaf disease, to add variation in the dataset, is used. Duplicate images were cleaned and removed. Images were preprocessed. For fair evaluation, the dataset was divided into training set (70%), validation set (15%), and testing set (15%). With 25 epochs, by employing Adam optimizer, model was trained. It is designed to extract visual features using four convolutional layers, five batch normalization layers, four pooling layers and two fully connected layers. To see how well trained model is performed, it is evaluated against some transfer learning models like MobileNetV2, ResNet50 and Vgg16 on the same dataset. Performance evaluation using accuracy, model size, time per image, and number of parameters showed that the proposed model achieved high accuracy and provided better discrimination between visually similar disease classes. Proposed model achieves 99.33% accuracy, has model size of 1.618 MB, 9.512 ms time per image, 424260 parameters. In contrast MobileNetV2 achieved 97.17% accuracy, has model size of 8.633 MB, 14.141 ms time per image and 2263108 parameters. ResNet50 achieved 44.93% accuracy, has model size of 90.01 MB, 64.661 ms time per image and 23595908 parameters. Vgg16 achieved 79.75% accuracy, has model size of 56.13 MB, 193.830 ms time per image and 14716740 parameters. Proposed model has performed well and doesn’t need numerous parameters, this shows that for practical implementation, it can be appropriate.