Computer Vision News - August 2019

Project Every month, Computer Vision News reviews a successful project. Our main purpose is to show how diverse image processing techniques contribute to solving technical challenges and real-world constraints. This month we review challenges and solutions in a Precise Agriculture project by RSIP Vision: Fruit sorting and grading. RSIP Vision has completed many projects in Precise Agriculture, some of them on fruit grading and sorting. A recent successful project is a good reason to share what we and our client learned that might be useful for similar projects. Read on! One of the recommendations that we give for this kind of projects is to take pictures of the same quality, so that all the parameters are under our control. This is why we prefer to be involved in the project as early as at the planning of the system. In this way, we can advice about the cameras’ positions, illumination, resolutions, image modality (IR or UV). The more we can influence the planning of system, the better our algorithm will perform. Oftentimes, the client is a producer of sorting machines, expert in the mechanics of its tool and in the conveyance of fruits within the sorting system. Besides perfectly mastering this part, the client is generally not able to automatically extract information from the fruit, especially in terms of specific fruit diseases and pests. Even when clients are able to design a system to sort fruits by size or color - where image processing is not too challenging – they need our help when the project requires to find defects and classify them. Client can do this by itself only when it employs a dedicated image processing expert team. In most cases, image processing is not the main expertise of the client: it is only one component of the service that it gives and it is generally recommended to call us to develop it, building on our long experience in this kind of projects. In recent projects, we have found that deep learning provides the best detection, when different defects need to be found: indeed, we want to verify not only the presence of defects, but also to classify them. For this purpose, you need a system which is able to receive photos as input, detect a range of defects and provide the needed information as output. When new kind of defects appear, the system can be quickly trained to detect them too, with no need to develop a specific algorithm for each defect. Annotation is one more challenging aspect of fruit sorting and grading, since the deep learning algorithm needs a large number of images, which in the real world the client does not generally have - at least not before fruit arrives to the packaging factory. Annotation is a challenging procedure, mainly because both the farmer and the manufacturer of the machine do not employ personnel in charge of labelling images. On our side, RSIP Vision has a trained ˆ 18