Performance Evaluations of Convolutional Neural Network (CNN)-Based Models for Semantic Segmentation of Plant Leaf Diseases

Abstract

Plant disease identification is important to sustain food production. Automated plant disease identification using Convolutional Neural Network has shown highly potential to provide effective solution to high accuracy and real-time plant disease detection. This paper presented the evaluations of five CNN-based models, namely DeepLabV3+ network with Resnet18/Resnet50/Resnet101, modified Alexnet, and Segnet with VGG-16 for semantic segmentation and identification of plant leaf diseases. The leaf images were acquired from Leaf Disease on Kaggle comprising four types of leaf diseases: bacteria, fungi, nematodes and virus. A total of 196 images were labeled for ground-truth development and training dataset. Image augmentation was conducted to increase the training dataset followed by assigning class weightage to the imbalanced classes. A total of 1,918 labeled images were produced and these images were used to train the five CNN-based models. All the pre-trained CNN-based models were modified to cater to the new leaf disease dataset and to optimize the semantic segmentation. The results showed that DeepLabV3+ network with ResNet-18 outperformed other models achieving 95.8% global accuracy for segmentation of the leaf diseases. This is followed by Segnet with VGG-16, ResNet-50, ResNet-101 and modified AlexNet.  However, upon closer study of the classes, the mean accuracy showed that AlexNet achieved better results compared to Segnet with VGG-16 and ResNet-50