Plant Pathology Disease Detection in Apple Leaves Using Deep Convolutional Neural Networks : Apple Leaves Disease Detection using EfficientNet and DenseNet

• Published in: 2021 5th International Conference on Computing Methodologies and Communication (ICCMC) pp. 1119 - 1127
• Main Authors: Srinidhi, V V, Sahay, Apoorva, Deeba, K.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 08.04.2021
• Subjects: Biological system modeling; Canny Edge Detection; Data Augmentation; Deep Convolutional Neural Networks; DenseNet; EfficientNet; Feature extraction; Feature Reuse; Image annotation; Image edge detection; Image segmentation; Machine Learning; Model Scaling; Pathology; Training
• DOI: 10.1109/ICCMC51019.2021.9418268

Over the years, many events of plant diseases have inflected suffering on untold millions of people worldwide by causing an estimated annual yield loss of 14% globally. Plant pathology is the science of plant diseases that attempts to improve the chances for survival of plants under unfavorable environmental conditions and parasitic microorganisms that cause disease. Temperature, pH, humidity and moisture are environmental factors contributing to development of plant diseases. Misdiagnosis can lead to misuse of chemicals causing economic loss, environmental imbalance and pollution and emergence of resistant pathogen strains. Current disease diagnosis is time consuming, expensive and based on human scouting. Automatic disease segmentation and diagnosis from plant leaf images can be reasonably useful than the existing one. Automatic plant disease detection involves image acquisition, pre-processing and segmentation, followed by augmentation, feature extraction and classification using models. This project uses Deep Convolutional Neural Networks models namely EfficientNet and DenseNet to detect Apple plant diseases from images of apple plant leaves and accurately classify them into 4 classes. The categories include "healthy", "scab", "rust and "multiple diseases". In this project, the apple leaf disease dataset is improved using data augmentation and image annotation techniques, namely Canny Edge Detection, Blurring and Flipping. Based on augmented dataset, models using EfficientNetB7 and DenseNet are proposed providing accuracy of 99.8% and 99.75% respectively and overcoming known shortcomings of convolutional neural networks.