Computer Vision News - August 2018

22 Computer Vision News Pallet Inspection by RSIP Vision 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 physical difficulties. This month we review RSIP Vision’s Deep Learning solution to inspect pallet status . RSIP Vision’s engineers can assist you in countless application fields. Project This is an interesting project, which we performed for a client who operates a big logistic center . He needs to perform a quality control on the pallets in current use. Manual inspection may not be cost effective and therefore we studied and implemented a deep learning based solution , designed to inspect each pallet and declare whether it is fit to be used again or not. Cameras being much cheaper now than in the past, an automated solution had all the chances to be cost effective. If once such an automated inspection system was justified only for expensive items like semiconductors , today it might be afforded even for inspecting rather inexpensive items like pallets. Cameras can see very well whether there are cracks in the pallet and what they are. Being the information thought of binary nature (pallet is reusable or not), it starts as a classification problem: in fact, damaged pallets might be fit for re-use depending on the position of the crack; if a pallet has a crack across its width, it cannot be used anymore and it must be disposed of. The client had already invested significant effort in developing a traditional image processing solution, however its precision and recall were not sufficient for use in the field. RSIP Vision introduced a Deep Learning solution able to provide much better results. But we did more: the system checks whether the crack extends through the entire width of the pallet or not. Via some manual annotation, we taught the neural network to check the crack and tell us for every pixel in the image whether it belongs to a crack and to which kind of crack (full cross- width or not), transforming the classification problem into a multi-class segmentation problem . This increased by an additional 5% the precision of the system, generating a significant saving for the logistic center. We adapted the network architecture to provide for each pixel its status within the entire pallet and not only local information. Even though we asked two outputs from the network (classification and segmentation), we were able to see that the additional work helped the training a lot and generated significantly better results. Its another example of how RSIP Vision’s deep learning algorithms were able to provide a much better solution to a real world problem. …transforming the classification problem into a multi-class segmentation problem.