A portable rice disease diagnosis tool based on bi-level color image thresholding

The rice plant is a cereal crop and staple food forpeople living in many Asian countries. Disease can lead to an underproduction of paddy-rice yield. In the past, diagnosing rice diseases required high-level skills to detect diseases accurately and was often very time consuming. Most mistakes happen due to human vision restrictions. In addition, each crop field disease has similar symptoms, making it hard to discriminate between each one. Additionally, current single thresholding methods are often not effective when separating color of lesion from the color of image background due to color information ignorance. To overcome these difficulties, we present a complete portable and real-time device for predicting rice diseases with an introduction of a bi-level thresholding method for segmenting regions of interest. The research methodology begins with camera calibration using distance versus a preset rectangular region of interest, image capturing via portable and real-time devices, image segmentation using a proposed bi-level threshold, system extraction based on image texture analysis and rice abnormality recognition via production rule method. Every captured image recognizes abnormality according to lesion shape and color. The rice abnormalities were classified into four classes including Blast Disease, Brown Spot Disease, Narrow Brown Spot Disease, and Sheath Blast Disease, while normal leaves are used as the test control. From all 120 real-time data images, only 85 data images are used for validation. Two experiments were conducted for optimal camera calibration testing via real time portable device and the proposed single threshold method testing. The experiment result shows that 20 cm is the optimal portable camera distance and that the proposed bi-level threshold achieves better accuracy (approximately 87%) compared with other state-of-the-art methods, namely the potential difference and color threshold methods (with approximately 82.4% and 54.2% accuracy, respectively).