Accuracy Assessment of Convolutional Neural Networks for Plant Disease Prediction with Multi-Stage Image Preprocessing

Abstract

Plant disease forecasting is among the advanced research areas targeted by most agricultural institutions and government bodies to enhance yields. Forecasting diseases is a very vital topic of international agriculture since crop health directly affects human health proportionately. Various plant organs, like root, stem, leaves, fruits, or a major proportion of the crop, are diagnosed to identify diseases, while various previous work employed images of plant leaves to identify disease. Previous works employed standard practices, optimization techniques, and machine learning methods for plant disease prediction but specifically for sets of leaf images. No work adopted a generalized approach to disease detection for any plant image data because the dataset was large, colored, aligned, and resolution varied. The goal of this paper is to adopt a Preprocessing Framework with a Convolution Neural Network (PF-CNN). It works on any plant leaf image. The preprocessing framework includes alignment, rotation, resizing, cropping, color transformation, and image enhancement. Final output images are passed through CNN for disease classification. As this work is the beginning of the research work, CNN is trained using pretrained images and checked against ground truth outputs to validate the disease class. Next, CNN is tested with test images and confirmed using validation images. Experiment output is cross-checked and compared with other analogous methods to analyse the performance of PF-CNN. Keywords: Preprocessing, CNN, Plant Disease Detection, Leaf Images, Tomato Leaves, Maize Leaves.