Plant Foliage Disease Diagnosis Using Light-Weight Efficient Sequential CNN Model

Abstract

The Precise and prompt identification of plant pathogens is essential to keep agricultural losses as low as possible. In recent time, deep convolution neural networks have seen an exponential growth in their use in phytopathology due to its capacity for rapid and precise disease identification. However, deep convolutional neural network needs a lot of processing power because of its intricate structure consisting of a large stack of layers and millions of trainable parameters which makes them inedquate for light computing devices. In this article, authors have introduced a novel light-weight sequential CNN architecture for the diagnosis of leaf diseases. The suggested CNN approach contains fewer layers and around 70% less attributes than pre-trained CNN-based approaches. For the experiments and performance evaluation, authors have chosen a benchmark public dataset consisting of 7012 images of tomato and potato leaves affected with early and late blight diseases. The performance of the proposed architecture is compared against three recent priorly trained CNN architectures such as ResNet-50, VGG-16 and MobileNet-V2. The average accuracy percentage reported by the proposed architecture is 98.02 and the time consumed in training is also much better than the existing priorly trained CNN architectures. The experimental findings clearly demonstrate that the suggested approach outperforms the recent existing trained CNN approaches and has a very less number of layers and parameters which significantly reduces the amount of computing resources and time to train the model which could be a better choice for mobile-based real-time plant disease diagnosis applications.