关键词: 玉米, 图像处理, 品种识别, 种子形态, 判别分析

Abstract: Seed identification plays a crucial role in seed quality testing and breeding programs in maize (Zea mays L.). Machine vision of seed surface features performances well based on a few experiments in maize. But the sample numbers in these studies were only 3–7 cultivars. To further examine the feasibility of image process application in discriminating numerous maize cultivars, six models were created and validated by means of principle component analysis and statistical discrimination analysis. The models comprised 4 categories or their combinations of 46 morphological traits extracted from scanned two-side images of 50 kernels each of 193 maize cultivars from Northeast and North China in recent years. Models of size, shape, texture, color, plus combination of latter 3 categories and combination of all 4 categories could correctly recognize cultivars at rates of 25%, 33%, 39%, 95%, 95%, and 95%, respectively, when cross-validated with all 9 650 kernels. Average refuse error rates were 90%, 90%, 86%, 45%, 47%, and 42%, respectively, and acceptance error ones were 92%, 92%, 88%, 46%, 48%, and 43%, respectively. These two error rates were highly and positively correlated between each other (r = 0.83**-0.91**). Machine vision wins the ad-vantages of low cost and high speed over manual or biochemical detecting methods, and is feasible to be applied to identification of numerous maize cultivars. The combination of shape, texture and color is the best model. Model performance may be promoted further with optimizing samples and structure.

Key words: Maize (Zea mays L.), Image process, Variety identification, Seed morpha, Discrimination analysis