Computer Vision News - March 2017

The kind of solutions presented above are still supplied and in use today. Company supplying sorting machines integrate visual systems and these algorithms. RSIP Vision does several steps forward with respect to these systems. Our expertise enables us to analyze and modify the algorithm, building on our experience in a very wide range of disciplines (medical, industrial and others); we also have knowledge and expertise in Deep Learning : using deep learning, it is possible to train the system to analyze the features of the inspected items over large samples of supervised images. After an initial classification, the network mimics our brain structure decision process: when it is loaded with images, it adjusts itself to inspect produce in order to grade it. One of the advantages of the deep learning technique is that we do not have to trouble ourselves with combinations of rules and complicate parameters. We simply equip the system with practical examples and define them as good or defective (or any other classification according to the features and the desired grades of quality). That is sufficient for the process to work automatically. The above is true not only for detecting defective products, but also for sorting them in different categories according to their features: shape, color, texture, tones and many other visible - and sometimes invisible - features. Sometimes, an internal spot which might potentially develop into a rotting factor is not visible from the outside, but is brought to our attention by a UV camera. An instance of notable external feature is whether a cucumber is straight or it makes an angle: it might be important to identify straight cucumbers, so that they can be profitably transported and sold in a market which values that kind of feature. A practical description of the process can be found in the below video, in which a set of oranges is classified and the system is trained to grade any orange that is presented to it. Needed hardware (specifically cameras) is no more than currently used cameras; of course, the system will benefit in case high resolution and better dynamic range cameras are used. For the learning phase, we prefer a powerful CPU, but regular off-the- shelf PCs might do the job as well. Regarding the choice between supervised and unsupervised process, standards exist today specifying what product is valid and what fails. A simple operator wouldn’t need to do anything more than applying the standard and classify produce according to the visual examples given. Read more on our website about this fruit grading project and other precise agriculture work by RSIP Vision . 21 Computer Vision News Project Fruit grading with Deep Learning Project Our latest video about fruits and vegetables grading with Deep Learning