Deep Learning Models for Wheat Diseases Detection and Recognition

Abstract

Wheat is a grass widely cultivated for its seed, a cereal that is a staple food around the world. However, cereal wheat is subject to many wheat diseases, including bacterial, viral and fungal diseases, as well as parasitic infestations. The need to use Deep Learning methods to identify automatically wheat diseases has become a challenge. In this paper, we proposed and compared two models based on Convolutional Neural Network (CNN) for wheat diseases detection and recognition. The convolutional layers of a CNN can be considered as matching filters derived directly from data images (images of healthy and unhealthy wheat). CNNs thus produce a hierarchy of visual representations optimized for our task. As a result of CNN training, a model is obtained-a set of weights and biases-which then responds to the specific task for which it was designed. One of the main strengths of CNNs is their ability to generalize, that is, the ability to process data never seen before. This allows a certain robustness to the heterogeneity of the background, to the image acquisition conditions and to the intra-class variability. A large image dataset of various wheat diseases, including healthy wheat, was used for training our models to learn, recognize and detect diseases and/or abnormalities in wheat.