Recognition of Mustard Plant Diseases Based on Improved Deep Convolutional Neural Networks

Abstract

Diseases of the mustard plant are a major threat to the quality production of mustard oil, but rapid recognition of these diseases becomes cumbersome due to the absence of expert identification infrastructures. This study introduced an improved diagnosis method for diseases of the mustard plant using deep convolutional neural networks (CNNs) to ensure sustainable improvement in mustard farming. First, the mustard plant dataset of nine classes is built using eleven image augmentation techniques that contain 47760 images of leaf, stem, and pod. Afterward, a CNN architecture, namely, MPNet, is designed and trained from scratch in this study that consists of deep separable convolutional layers and inception modules, which realize 97.11% accuracy in recognizing 2388 test images. The recognition performance of MPN et is also compared with four state-of-the-art CNNs, where MobileNetV2 acquired 92.83% test accuracy. The results authenticate that the proposed MPNet can competently recognize diseases of the mustard plant.