April 2020

n MACHINE VISION On production lines, rule-based machine vision systems can inspect hundreds, or even thousands, of parts per minute with a high level of accuracy. They are more cost- effective than human inspectors. Their output is based on a programmatic, rule- based approach to solving inspection problems. Such traditional rule-based machine vision systems are ideal for: guidance (positioning and orientation); identification (barcodes, data-matrix codes, marks and characters); gauging (comparing distances with specified values; and inspection (flaws and other problems such as missing seals or broken parts). Rule-based machine vision is great when used with a known set of variables: Is a part present or missing? How far is this object from that one? Where does the robot need to pick up this part? These jobs are easy to deploy on assembly lines in a controlled environments. But what happens when things aren’t quite so clear-cut? This is where deep learning comes in. It can be used to: n solve vision applications that are too difficult to program using rule-based algorithms; n handle confusing backgrounds, and variations in the appearance of parts; n maintain factory-floor applications and re- train them with new image data; or n adapt to new examples without needing to reprogram core networks. A typical application is to look for scratches on the screens of devices such as mobile phones and tablets. Such defects can differ in size and location, or across screens with different backgrounds. Deep learning takes such variations into account and differentiates between good and bad parts. Plus, training the network on a new target (such as a different screen) is as easy as taking a new set of reference pictures. Traditional rule-based machine vision systems face serious challenges when inspecting similar parts with complex surface textures and variations in appearance. They will almost always reject “functional” faults, but not “cosmetic” anomalies. It is difficult for traditional vision systems to distinguish between such defects. Some types of defect are notoriously difficult for traditional machine vision systems to program and solve because of multiple variables – such as variations in lighting, or changes in colour, curvature or field of view – that can be hard to isolate. Here again, deep learning brings appropriate tools. In short, traditional machine vision systems perform reliably with consistent and well-manufactured parts. For the complex situations that need human-like vision with the speed and reliability of a computer, deep learning is a game-changing option, but might entail bigger hardware investments because it needs more processing power and storage capabilities. Rule-based machine vision and deep- learning-based image analysis are complementary to each other, rather than being an either/or choice for next-generation automation tools. In some applications, such as measurement, rule-based machine vision is still the preferred and more cost-effective choice. For complex inspections, involving wide deviations and unpredictable defects – too numerous and complicated to program and maintain within a traditional machine vision system – deep-learning-based tools offer an attractive alternative. n • Light and extreme heavy duty models • ATEX certified models for Gas, Dust andMining • Stainless steel for optimal corrosion resistance • Large hollow and solid shaft designs • Radial / Axial cable or plug connectivity • Protection class up to IP69K • Extensive range of Fieldbus protocols www.impulseautoma琀on.co.uk | tel: +44 (0)1264 364194 | ISO 9001 registered company Absolute and Incremental Rotary Encoders for Speed, Direc琀on and Posi琀onal Feedback UK O cial Distributor of Hengstler Traditional machine vision systems and those based on deep learning each have their own strengths and weaknesses, and are complementary to each other.