Rice Leaf Diseases Classification Using Discriminative Fine Tuning and CLR on EfficientNet

Abstract

Rice cultivation in Nepal is effect by many factors, one of the main factor is rice leaf diseases which limits the crops production. Image classification of rice leaf classify different rice leaf diseases. Image dataset of rice leaf diseases is taken from open source platform. Pre-processing of image is done which is followed by feature extraction and classification of images. This thesis presents image classification of rice leaf diseases into four classes, namely: Brown Spot, Healthy, Hispa, Leaf Blast using Convolutional Neural Network (CNN) architecture EfficicentNet-B0 and EfficicentNet-B3 based on fine-tuning with cyclical learning rate and based on discriminative fine-tuning. It is found that the test accuracy of EfficientNet-B0 is 81.96% and EfficientNet-B3 is 85.12% based on fine-tuning with cyclical learning rate and the test accuracy of EfficienNet-B0 is 83.99% and EfficientNet-B3 is 89.18% based on discriminative fine-tuning for 15 epochs. The results also conclude that the CNN architectures work better with discriminative fine-tuning than on fine-tuning with cyclical learning rate. The classification models of EfficicentNet-B0 and EfficicentNet-B3 are evaluated by recall, precision and F1-score metrics.