Mills and silos use strict quality criteria to determine whether to accept or reject deliveries of cereal grains. However, the Besatz analysis – which determines the composition of the cereal – is still performed manually. The “QualiKorn” research project, part of the InnoNet program run by the German Federal Ministry of Economics and Technology, aims to automate this process. The project team includes a variety of research institutes, industrial enterprises and industry associations, including the Steinbeis Transfer Center for Quality Assurance and Image Processing in Ilmenau. The center provides the project with innovative new components for automated analysis, including cameras, controllers and light sources.
In manual Besatz analysis, a sample of the cereal grain delivery is taken and sieved, roughly separating the contents into different fractions according to grain size. A laboratory worker then inspects these fractions visually and removes impurities manually. The results of this Besatz analysis help to decide whether to accept or reject the cereal delivery and how much to pay for it.
Automatic Besatz analysis is based on the principle of recording images of the individual grains and impurities as they fall, using a charge-coupled device (CCD) color line-scan camera. Line-scan cameras can record images non-stop at very high resolutions, making them ideally suited to a constant stream of free-falling grains. As the grains are constantly in motion, the second dimension to the image is created automatically. When recording color images, color line-scan cameras split the visible spectrum into three color channels: red, green and blue (RGB). To achieve optimum color resolution and dynamics, the device uses a three-line camera with a beam splitter to direct the light onto the three separate lines.
This setup requires a combination of three light sources – two to act as incident light and one as transmitted light. A short exposure time avoids motion blur as the grains fall, but this requires high light intensity. So the light source must have extremely high output, excellent homogeneity and clearly defined spectral emissions – as well as high efficiency, optimum durability and a compact size. To meet these requirements as best possible, the project used state-of-the-art LED technology together with specially adapted cooling equipment and a new type of projection lens.
When it comes to detecting individual objects in the grain sample, imaging is merely the first step in the process. This is followed by segmentation of the images, which isolates the objects of interest (in this case, grains and impurities) and separates them from the image's blank background. This reduces the volume of data by around 95 per cent, as all irrelevant areas are deleted and not processed further. Next comes feature extraction, which determines the individual features of the photographed objects. This involves assessing the color, shape and texture of the grains and impurities to differentiate them from one another. Combining these different factors results in a feature vector made up of around 200 feature values. Support vector machine (SVM) methods are used to classify the photographed objects based on their feature vectors.
Using a data record consisting of 23 object classes, the project achieved detection rates of 81–99 per cent for the four main combined classes at a sample throughput of 50 g/minute. The system delivers images of all objects, the weight of the sample as a whole, the proportion of usable wheat and statistics on the sample's composition. This data can then be entered into forms provided by the companies (in this case, mills and silos) and saved. The company designlab-weimar has developed a provisional concept study for using the device once the project has been completed.
Dr.-Ing. Peter Brückner | Katharina Anding | Martin Dambon | Daniel Garten
Steinbeis Transfer Center for Quality Assurance and Image Processing (Ilmenau)