Classification of capsicum leaf disease from a complex cluster of leaves using an improved multiple layers ShuffleNet CNN model

• Main Authors: Chyntia Jaby, Entuni, Tengku Mohd Afendi, Zulcaffle, Kismet, Hong Ping
• Format: Article
• Language: English
• Published: ACCENTS 2023
• Subjects: T Technology (General); TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://ir.unimas.my/id/eprint/42155/3/Classification.pdf; http://ir.unimas.my/id/eprint/42155/; https://accentsjournals.org/paperInfo.php?journalPaperId=1536&countPaper=232; http://dx.doi.org/10.19101/IJATEE.2022.10100509

Capsicum, also known as chili pepper or bell pepper, is cultivated worldwide and holds significant economic importance as a condiment, vegetable, and medicinal plant. One of the major challenges in capsicum cultivation is the accurate identification of leaf diseases. Leaf diseases can have a detrimental effect on the quality of capsicum production, leading to substantial losses for farmers. Several machine learning (ML) algorithms and convolutional neural network (CNN) models have been developed to classify capsicum leaf diseases under controlled conditions, where leaves are uniform and backgrounds are uncomplicated. These models have achieved an average accuracy of classification. However, classifying diseases becomes relatively challenging when a diseased leaf grows alongside a cluster of other leaves. Having a reliable model that can accurately classify capsicum leaf diseases within a cluster of leaves would greatly benefit farmers. Therefore, the aim of this study was to propose a model capable of classifying capsicum leaf diseases both from a uniform background and within a complex cluster of leaves. Firstly, a dataset comprising images of diseased capsicum leaves, including discolored leaves, grey spots, and leaf curling, was acquired. Subsequently, an improved multiple-layer ShuffleNet CNN model was employed to classify the different types of capsicum leaf diseases. The proposed model demonstrated superior performance compared to existing models, achieving a classification accuracy of 99.30%. Furthermore, it was concluded that augmenting the layers of ShuffleNet, utilizing a 0.01 initial learning rate, employing 50 maximum epochs, using a minibatch size of 64, conducting 10 iterations, and incorporating 205 validation iterations all contributed to the improved ShuffleNet model's success.