基于机器视觉图像特征参数的马铃薯质量和形状分级方法

马铃薯自动分级过程中,存在既要保证分级精度又对分级速度有一定要求的难点问题。该文探讨了利用机器视觉技术快速获取马铃薯图像特征参数,结合多元线性回归方法,建立马铃薯质量和形状分级预测模型,实现基于无损检测的马铃薯自动分级。搭建了同时获取马铃薯三面投影图像的机器视觉系统,通过图像数据处理获得马铃薯俯视图像轮廓面积、两侧面图像轮廓面积、俯视及侧面图像外接矩形长度及宽度数据等图像特征参数,通过多元数据回归分析,建立了马铃薯质量和形状分级预测模型。选择100个试验样本运用该方法进行质量和形状分级模型构建和预测,采用电子称获取样本实际质量,采用目测法对马铃薯进行形状分选。对比试验结果表明,质量分级相关度系数R为0.991,形状分级分辨率为86.7%。表明该方法对马铃薯质量和形状分级进行预测具有可行性,可运用于马铃薯自动分选系统中。

Potato grading method of weight and shape based on imaging characteristics parameters in machine vision system

Abstract: Potato is cultivated as a major food resource in China. Manual grading is labor intensive. Machine vision system is one of the modern grading techniques and is becoming research focus. Weight and shape of potato are important indexes to divide potato grade. Generally, weight and shape of potato have significant positive correlation with outside dimension parameters of potatoes. It is the key to increase potato grading accuracy and speed in order to quickly obtain the imaging feature data possessing high correlation with potato weight and shape and to establish a strong correlation predictions estimation model for potato weight and shape. The focus of this research was to develop a potato grading method of weight and shape by means of image processing in the machine vision system. Firstly, the machine vision system was established, which can capture a potato's three projection images simultaneously using a V-shaped plane mirror. One hundred potato samples were randomly selected, which were constituted of large, medium, small sizes, approximation sphere and approximation ellipsoidal according to artificial visual determination. Then the image feature parameters were obtained employing the digital image processing technology, including the contour areas in top view and two side views, the length and width of circumscribed rectangle in projection image of every potato sample. Secondly, the feature parameters with high weights value were selected using PCA (Principle Component Analysis) method in Unscramble software. The analysis results showed that the first two principal components explained 96% information contained in all characteristic data, and the scores of 100 potato samples were distributed in obvious three regions in the score graph with small size located the lower-left area, medium size located the middle area, and large size located the upper-right area. The predicted model of potato weight was constructed by means of multiple linear regression analysis using data of three contour areas in top view and two side views of every potato sample. The actual weight of the potato samples were gained by an electronic balance, and the correlation coefficient was 0.991. The distinguish accuracy were respective 90%, 100%, 90% for large, medium and small sizes in potato sample test set. Finally, potato shapes were analyzed by PCA using feature data, including the length and width of circumscribed rectangle in three projection image of potato samples. The score graph showed that the first two principal components explained 95% information contained in all feature data. The feature data scores were used to divide 100 potato samples into two types. In order to use the image characteristic parameters to determine the shape of potatoes, we set two dummy variables as -1 and 1, respectively, which represented approximation sphere and approximation ellipsoidal. The prediction model of potato shape was then established by the partial least squares discriminate analysis. The actual shape of the potato samples were decided by artificial ocular measurement, with a ratio of classification of 86.7%. Grading test of shape classification was completed for 40 potato samples in test set using the regression equation. A potato sample with a positive calculated value was judged as approximation ellipsoid, and a potato sample with a negative calculated value was judged as approximation spherical. Grading accuracies for approximation ellipsoid and approximation spherical were 83.3% and 89.3%, respectively. Our research indicated that the regression model for shape grading was reliable. Therefore the approach for non-detection inspecting potato weight and shape were effective and feasible, which can be applied in a potato grading system.