Investigation on novel convolutional neural network architectures for classification of diseases in plant leaves

Abstract

Plant disease is one of the major causes for an overall decrease in newlineyield and reduces the profit from the farming segment. Convolutional Neural newlineNetworks (CNN) has the potential to classify multiple classes with automatic newlinefeature extraction capability and can be light weight to run on edge devices. newlineTherefore, this research work concerns with development of novel lightweight newlineCNN architectures for leaf disease classification. Thesis consists of five newlinecontributing works for recognizing leaf diseases in plants. newlineTo acquire significant features in lesser number of trainable newlineparameters, a novel microarchitecture called triconv module is proposed. newlineThese modules are arranged linearly in three branches to obtain colour, shape newlineand texture features and the developed CNN architecture is called quotTriconv newlinenetwork Version Iquot. newlineFurther, to introduce the multiresolution and spectral features, a newlinenovel hybrid CNN architecture called triconv network version II is proposed newlineby combining features from wavelet decomposition with CNN feature maps. newlineA novel CNN microarchitecture called delta blocks is proposed to newlineovercome the drawbacks of the bottleneck approach and processes feature newlinemaps in parallel. These are stacked together to form a novel CNN architecture newlinecalled Delta Tributary Network . Further, the proposed delta block is built newlineusing a novel bifurcation layer that splits feature maps into partitions and newlinecontrols the number of input channels given to 3and#61620;3, depthwise and 1and#61620;1 newlineconvolution layers. newline