Deep Bi-linear Convolution Neural Network for Plant Disease Identification and Classification

Abstract

Plant diseases that occur through various sources is a threat which causes a huge loss in production of farming. Detection of infected plants is necessary to take required preventive measures to protect the plant and preserve it. As manual identification of the disease requires expertise to determine the pathological status of the plant regarding the infection levels. So, automated identification is preferred by driving human-level intelligence to the machine using deep-vision approaches to provide fast, reliable and accurate solutions.In this research, automated identification and classification of 9 distinct plant-leaves with healthy and infection classes is developed using Bi-Linear Convolution Neural Network (Bi-CNN’s). This neural architecture is constructed inspiring from the visual perception of the human brain with two cortical pathways. The hyperparameters are fine-tuned by the scheduling training procedure to attain faster convergence. The model is generalized on a triad of testing splits ranging from 10 to 50%. This model is evaluated on various standard classification metrics and when testing samples are increased by 5x (i.e. from 10 to 50%) the deviation in the accuracy score is very minute (0.27%) which resembles the resilience to unseen samples. The AUC obtained for all the models for variant test samples is at least 99.92%.KeywordsBilinear CNN’sPlant automated diagnosisComputer visionFine-tuning CNN’sNeural networksPlant leaf classification