Automatic Grading of Green Intensity in Soybean Seeds

Abstract

In this work we introduce a low cost machine vision system for grading problems in agriculture. Instead of a careful evaluation of a given quantity over a reduced number of samples with a high cost dedicated equipment, we propose to measure the quantity with less precision but over a much bigger number of samples. The advantage of our procedure is that very low cost vision equipment can be used in this case. For example, we used a standard flatbed scanner as an integrated illumination plus acquisition hardware. Our system is aimed at the quantification of the amount of chlorophyll present in a production batch of soybean seeds. To this end we arbitrarily divided green seeds in four classes, with a decreasing amount of green pigment in each class. In particular, in this work we evaluate the possibility of an accurate discrimination among the four classes of green seeds using machine vision methods. We show that morphological features have low discrimination capabilities, and that a set of simple features measured over color distributions provides good separation among grades. Also, most errors are assignations to neighbor grades, which have a lower cost in grading. The good results are almost independent from the classifier being use, Random forest or Support Vector Machines with a Gaussian kernel in our case.