Leaf Spot Disease Image Classification for Groundnut Crop using Deep Convolutional Neural Network

Abstract

Groundnut is an important cash crop cultivated in over 100 countries across the world and India is one of the largest
countries in the world producing groundnut with an average of 745 kg/ha. The groundnut crop is prone to infection by
pathogens and viruses that cause disease in the leaf, stem, and root which affects the yield. The Convolutional Neural Network
(CNN) used for image classification takes a long time to build the neural network when the input images are in high resolution.
CNN also have difficulties in identifying the patterns when the images have different background and is there any tilt or degree
of rotation present in the input image. To overcome these difficulties, this paper proposed a Convolutional Neural Network
Algorithm for Groundnut Disease Prediction (CNN-GDP) that is designed to diagnose early and late leaf spot disease in
groundnut crops. The input image is properly pre-processed and compressed with image compression techniques such as
Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT), and Discrete Wavelet Transform (DWT). To
study the performance of the CNN with different input image formats the collected input image is enhanced with
Thresholding Operations, Homomorphic Filters, and Contrast Stretching techniques. The database is reconstructed with
the output images of each image processing algorithm. Each database is trained with the CNN which is built using the ResNet-
50 Architecture and the performance of the CNN for each analyzing the transformed database is measured with the
performance metrics. The results show that the CNN trained using DFT and DCT classifies the leaf spot infected leaves
correctly with higher accuracy and a low error rate. The computation time taken to build the Convolutional Neural Network
with ResNet-50 architecture using the transformed image is extremely reduced when compared with the existing CNN.