Research On Features Extraction And Varieties Identification Of Rice Seeds Based On Machine Vision

Rice is one of the most important foodcrops in China; the country has more than 60% of the rice-eating population. The yield and quality of rice is directly related to the quality of rice seed. At present, the identification methods of rice seed variety purity include phenol dyeing, seedling identification and molecular marker method, but these methods can not meet the requirements of rapid varieties identification because of the high expense and long inspection cycle. In view of this problem, the paper selects nine rice seed varieties as the objects for study, and researches feature parameters automatic extraction and varieties rapid identification by using machine vision technology and multivariate statistical method.The main content and conclusions of this research were as follows:(1) The machine vision system for rice seeds variety identification was established, which was composed of the lamp-house, the light box, the digital camera and the computer. The whole system was portable and practical, and the good original images could be taken by it.(2) The original images were pretreated, such as color image gray processing, image enhancement, image segmentation, image denoising, edge detection etc. Compared with the results of other pretreatment algorithms, the most suitable image processing for this research were illustrated as follow: the media filtering of 3x3 square window was used to enhance image, the optimal threshold method was used to achieve image segmentation, the open and close operation of the morphological operation was used to remove noise, and the edge of rice seed detected by Roberts operator was the best.(3) The algorithms of 13 feature parameters extraction were proposed, including rice seed region area, perimeter, length, width, the maximum radius, the smallest radius etc. The program of algorithms was made to extract features automatically. Experiments showed that the algorithms of feature extraction were efficient and accurate.(4)The categories model was established by Bayes discriminance, three categories of rice seeds included Indica rice, Glutinous rice, Japonica rice. Discriminance testing showed that discriminant ability of the categories model was obvious; features of the largest contribution to category identification were width, the smallest radius and one-fourth width; the correct discrimination rates of the training set of three categories were 99.7%,78.3%,98.7%. The varieties model was established by Bayes discriminance, nine varieties of rice seeds included IRBB10, Zhenzhuai, Peiai64, Yifangnuo, Haonuoliang, Gaogandanuo, Zhonghua11, Balilla, Qiuguang. Discriminance testing showed that discriminant ability of the varieties model was obvious; features of the largest contribution to variety identification were length-width ratio, radius ratio and width; the correct discrimination rates of the training set of nine varieties were 97%,100%,99%,76%,64%,63%,66%,99%,69%. The models were tested by using new samples, the identification accurary of categories model was 90%, the identification accurary of varieties model was 76.7%.(5)A software system for rice seed feature extraction and variety identification was developed. The functions of software system included image pretreatment, feature extraction and storage, variety identification etc., friendly interface, stable operation and easy maintenance were the prominent advantages of this software system.