Computer Vision News - April 2018

21 Application: Aquifi Computer Vision News Carlo says that it is always possible to find something that works for a limited set of problems, so generalization is a big issue. On the other hand, for the identification part, they rely on deep learning and convolutional neural networks. Aquifi developed a set of 3D convolutional neural networks that are very reliable and allows for generalization. The use of 3D for deep learning is something they are pursuing. It’s not a trivial task, but it has got major advantages in that it allows reducing the amount of training samples for training your algorithm and it allows for better recognition performances. Carlo tells us where he would like the solution to go next: “ We want to start using the operator more, so we no longer need a training set. Right now, we install our system on the line, collect data for a few days, and then there is a big effort in labelling it. If we can completely remove this by using unsupervised techniques and some help from the operators on the line, just observing the operators, that’s a major benefit. What we can do always is to incrementally build on top of what we currently have. I don’t know how much we’ll be able to push to have this completely automated solution, but we are going to substantially decrease the amount of required labelling by exploiting reinforcement learning, transfer learning and active learning techniques. That’s something we can definitely do. ” Carlo concludes by telling us what he is most proud of about Aquifi: “ We solve problems that so far we haven’t seen other people take on. We are dealing with defects or anomalies that you cannot really model, because every defect is different. We leverage 3D and deep learning for tackling these problems and we are really proud of the techniques we’ve developed so far. ” This article was first published three weeks ago on our magazine Bay Vision, spring 2018 “We solve problems that so far we haven’t seen other people take on. We are dealing with defects or anomalies that you cannot really model, because every defect is different. ” Application