Optical sorters, like the OptoSelector OS 901 or the optical fine seed sorter OS f-class, have to fulfil what had been done in the past by humans and in some cases, is still done by humans, i.e., detect and pick objects out of a material stream that do not belong to it. In the case of seed processing objects like foreign seeds or fusarium infested seeds, foreign artificial and non-artificial objects have to be detected and have to be safely ejected. The human brain and the human eye have undergone an amazing evolutionary process which allows deciding intuitively and easily what is good and what is bad. An optical sorter is in the first place a “stupid” machine. Its eye, the full colour camera, absorbs photons which produce an electrical voltage signal and is represented as a digital number in the computer. The image which we see on the computer is a matrix of numbers that has to be linked to higher level, abstract features and labels, like colour, size, good object, bad object. Machine learning is the link between a cloud of numbers and our world of perception.
When a new type of seeds, e.g. corn, shall be sorted with the OS 901, the user has to perform a teach-in process in order to let the machine know what has to be accepted and what type of seeds have to be rejected. Therefore, the user has to take images of corn kernels that represent either only good kernels or it represents good kernels and bad kernels and has to label the corn kernels accordingly, which serves as a basis for the machine learning process. Quite a few machine learning algorithms exists, like support vector machines, genetic algorithms, Bayesian networks, decision tree learning, deep learning etc. All of them have one thing in common: none of them work perfectly for such a wide natural spectrum of tasks like optical seed sorting. It gets even more complicated when taking into account the different requirements of customers for the same seed type.
The classification result of the machine learning algorithm of the OS 901 covers meanwhile a large amount of seed types and requirements, which makes life so much easier for the operator. To also cover the cases if a type of defect is not detected properly by the OS 901 machine learning algorithm, the operator can in addition define manually a classification, using a broad range of colour and geometrical features. The combination of the preferred easy automatic teach-in way of learning and doing adjustments manually, ensures applying the OS 901 and the OS f-class to a very broad range of seeds.