Detection of Defects in Rice Seeds Using Machine Vision

Three image-processing algorithms were developed to detect external defects of rice seeds such as germ, disease, and incompletely closed glumes. The rice seeds used for this study involved five varieties: Jinyou402, Shanyou10, Zhongyou207, Jiayou, and IIyou. Images of the samples with both black and white backgrounds were acquired with a color machine vision system. Each original image was preprocessed to create a mask for the seed region. For judging the presence of germ, 16 contour features were extracted and analyzed using principal components analysis. In addition to this, four back-propagation neural networks were created and trained with typical data sets of the four varieties. The algorithm developed for recognition of germ achieved an average accuracy of 99.4% for normal seeds and 91.9% for germinated seeds on panicle. The mean hue value and its deviation of the seed region determined with a block method were extracted as features of disease recognition. The corresponding algorithm developed for inspecting diseased seeds based on color features achieved an accuracy of 92.1% for normal seeds, 94.8% for spot-diseased seeds, and 91.1% for severely diseased seeds. Using radon transform, the group number of post-processing images proved to be a good indicator of incompletely closed glumes. The relevant algorithm was developed and achieved an accuracy of 98.6% for normal seeds, 98.6% for seeds with fine fissures, and 99.2% for seeds with unclosed glumes. The results showed that the three algorithms achieved desired accuracy.