Xception model for disease detection in rice plant

Abstract

Rice is a staple meal that helps people worldwide access sufficient food. However, this crop has several illnesses, significantly lowering its production and quality. Because of this, it is imperative to conduct early disease detection to halt the spread of infections. Because of this, it is desirable to develop an automatic system that will help agronomists, pathologists, and indeed growers in directly diagnosing rice diseases. This would allow for preventative measures to be done as quickly as feasible. In this day and age of artificial intelligence, researchers have experimented with various learning approaches to discover diseases that can affect rice plants. Deep learning has recently seen considerable use in many computer vision and image analysis fields, becoming one of the most prominent machine learning algorithms. Deep learning has also recently found substantial usage in many computer vision and picture analysis fields. On the other hand, deep learning methods have seen very little application in plant disease recognition, except for some ongoing research centered on the problem and using a public dataset of pictures magnified to show plant leaves. Because of their high computational complexity, which requires a huge memory cost, and the complexity of experimental materials’ backgrounds, which makes it difficult to train an efficient model, deep learning methods have only seen limited use in plant disease recognition. This is due to several factors, including the following: The Inception module was improved to recognise and detect rice plant illnesses in this research by substituting the original convolutions with architecture based on modified-Xception (M-Xception). In addition, ResNet extracts features by prioritising logarithm calculations over softmax calculations to get more consistent classification outcomes. The model’s training utilised a two-stage transfer learning process to produce an effective model. The results of the experiments reveal that the suggested approach can achieve the specified level of performance, with an average recognition fineness of 99.73% on the public dataset and 98.05% on the domestic dataset, respectively. Our proposed work is better as per existing methods and models.