Forgotten pens and needles are among the costliest challenges in the laundry industry worldwide. Pens destroy a lot of laundry while needles pose a major security risk to employees. X-ray technology is one part of the solution, especially when the machines are able to analyze the images on their own.
At Inwatec, a team of programmers is working hard every day to solve problems that one can literally find in the pockets of the clothes that pass through the washing process. One of these people is Tudor Morar, a young Romanian-born robot engineer who joined the team when he invented the basis for the solution that is now running on all of Inwatec’s X-ray machines during a single weekend in 2015.
“Normally you would make the X-ray system search for dark spots in the pictures but that doesn’t work well enough with plastic parts. Therefore, we have developed our own algorithms. I can’t reveal business secrets but I can say that the novelty resides mainly in the fact that we consider the shape of objects instead of their material,” says Tudor who gets very excited when speaking about the many challenges of analyzing garments that he and his colleagues face.
“Garments are a big challenge because you won’t find two X-ray images that are alike. The seams always look differently and there are all kinds of buttons and zippers one has to take into account. The current technology is really good when it comes to sorting garments such as hospital clothes. However, with heavy fabrics, there comes the challenge. With these, the machines make errors too often. While the machine is supposed to let zippers and large buttons pass, it is expected to reject screws and small nuts,” says Tudor.
Massive computing power available
One part of the solution to the problem are better algorithms that are able to spot the foreign elements. Another important aspect is artificial intelligence (AI) that can help the human operators with making the right choices.
“One of the biggest challenges regarding the X-ray sorting is that it cannot look for foreign elements too thoroughly. If we make it too sensitive, clothes with many creases get rejected. On the other hand, if it is not sensitive enough, needles and pens slip through and end up in the washing,” explains Tudor Morar who has written all of Inwatec’s programs from the scratch and therefore knows how to assist the computers during the process.
“The solution is to teach the computer how to spot foreign objects on their own. Therefore, we started using AI. We have just started to rent some Amazon servers optimized for computations in order to be able to work with larger amounts of data. To give a rough estimate, the processor cores of Amazon’s servers are 64-times more powerful than our computers, but it is mainly the graphics processing unit that is much more powerful than anything we had available before,” explains the Romanian programmer. He and his team are, however, not even closely challenging the full capacities of these supercomputers.
“Fortunately, we have a great collaboration with our customers who are, of course, also interested in continuously improving our software. That’s why we have received plenty of real images from our customers in Norway, Germany, and the Netherlands. I think we are talking about 15,000 images that we have to analyze. Theoretically, the machine could sort one million images a day but there is an important obstacle: We have to sort the images first into two groups, reject or pass, which has to be done manually,” says Tudor Morar.
Computers have to learn how to think by themselves
At Inwatec, we have better jobs for the programmers than using their working time for manually feeding computers with images. For this purpose, Tudor Morar and his colleagues embarked upon a very ambitious plan:
“We are using deep learning algorithms that are designed to function like neurons in the human brain. That way, the computer is able to learn by itself without anybody having to take care of it during the whole time,” says Tudor who together with programmer Martin Bulin already are in full swing of training his electronic ‘students’.
“The AI needs a fairly precise description of what it should look for. The more we can specify it, the better the results will be. Not everything works equally well but there are patterns in the images that can be detected and then be used in the deep learning process in order to improve the results. This is what we are very interested in when teaching the X-ray machines how to work,” says Tudor.