Rapid image detection and recognition of rice false smut based on mobile smart devices with anti-light features from cloud database

Abstract

Using deep learning-based static image processing technology to identify crop diseases has become an important topic in recent years. However, this technology usually requires a good network environment and powerful computing equipment. It is not conducive for farmers to identify rice diseases on-site in under-developed areas with poor network signals. To address this problem, a crop disease mobile identification system that can adapt to a poor network environment is proposed. Rice morphological characteristics are used for support vector machine (SVM) model to realise offline recognition of rice false smut (RFS) analyzed by histogram of oriented gradient (HOG), circumscribed rectangle aspect ratio (CRAR) features and tilt correction algorithms. Images of rice lesions were obtained through field shooting. These images were used to build a database to train a cloud convolutional neural networks (CNN) recognition model to correct the offline recognition results when the device is in a poor network environment. This database can improve the lack of adaptability of ordinary public databases in the field environment. Moreover, this system is compressed into smart phones to facilitate on-site identification by farmers. This system has a 98% recognition rate of RFS and a recognition speed of 4s. In the case of low specification equipment configuration and a poor field network environment, this system is superior to other recent feature extraction methods. • Establishing a cloud database to make up for the poor performance of most existing database. • Both online and offline working modes. • The system is compressed into smart phones to facilitate on-site identification by farmers. • The parameters CRAR and HOG are used to reduce the influence of light on recognition results.