Plant Leaves Disease Classification Using Bayesian Regularization Back propagation Deep Neural Network

Abstract

Abstract Automatic leaf disease segmentation and detection from plant images has recently become a major research area around the world. The proposed method uses plant images to automatically segment and classify different leaf disease regions. The proposed procedure consists of four steps: Pre-processing is used in the first step to reduce the amount of background noise in the plant image using the Wiener filter. The disease spot is then detected using the hue histogram on the HIS model and these disease spots are then segmented using the K-means algorithm applied on the L*a*b* colour model and highest hue value calculation on HSV colour model. Finally, seventeen colour and texture features are extracted from the disease segment and these features are fed to a forward-propagation deep neural network (FPDNN) classifier which classifies the diseases. We have used the Bayesian regularization back propagation algorithm to fine-tune the results. We have applied FPDNN on varying hidden layers ranging from 1 to 40 and achieved highest accuracy i.e. 97.18% at 19 hidden layers which is larger than other state of art classifiers. Proposed method is implemented in matlab 14a.