A vision of the future

A UK company has developed an intelligent vision system that detects structural surface and colour defects to whatever level the user requires. Dean Palmer reports A next-generation vision system has been developed that is able to find structural surface defects and colour faults to whatever level customers require for their manufacturing process. The 'webSpector' imaging machine, developed by Shelton Vision Systems based in Leicester, has an in-built self-learning system that is capable of teaching itself to find errors and reporting them to operators or production managers. And, the system's innovative 'webCorder' enables the entire set up test run to be played back, which, according to Shelton's director Mark Shelton, results in "faster set up times and reduced scrap." So how does it work? Mark Shelton explained: "It uses a combination of powerful, off-the-shelf server grade PC technology with flexible software architecture and top-end digital imaging technology, to provide users with a scalable system for finding surface and colour defects. All the hardware components are replaceable and upgradeable. The system can even run in harsh environments." WebSpector is a potential solution to a common problem faced by manufacturers of high volume, highly diverse products who also have high fault detection requirements. Finding a vision system that can actually cope with this type of production process is difficult. Mark Shelton explained: "Many companies require high sensitivity for defects with lower contrast than the surrounding texture. The fabrics are of high quality, with detection accuracy and low false alarm rates critical to maintaining production yield. "Other manufacturers have a very large defect range of different sizes, shapes, orientations and contrasts. One customer we have has more than 1,500 different products, with new products and styles being introduced frequently. The other issue you have with installing a vision system in-process, is that non-permanent defects like features can occur on the material which need to be ignored. You don't want to stop the production line for these." WebSpector is built on PC-based processing modules because of their rapid development and powerful Pentium 4 MMX/SSE instruction sets, which are very suitable for 8-bit image processing. Normally, there is a central PC that interfaces to the factory and which controls all the other PCs that work in the background and are connected to one or more line-scan CCD cameras. The central PC also performs defect classification, report presentation and other user interface tasks. There are three camera imaging planes, one lit from above the material or fabric, one from behind and one at the low angle to the top surface. The system combines the information from the three imaging planes to help it make decisions as to what each defect type is. Shelton developed the original webSpector system in 1999, primarily for the woven and non-woven textiles industry, where the company's roots were. But the latest (Mark 3) system is equally suitable for inspecting aluminium, steel, plastic, rubber, laminated materials, composites, coated materials and technical textile surfaces. The system can also detect the relative change between two colours to better than one delta E (ie. better than the human eye). Shelton's system approach to solving the "high product range/high defect sensitivity" problem is patented and involves the combination of three separate ideas into the overall vision system. These are the 'webCorder', the 'webGrouper' and the 'webTrainer'. Shelton's development engineer Dr Millman explained: "With such a large range of styles, it is impractical to carefully train each one to the necessary standard. Hence, the styles need grouping to a controlled extent. This is the function of the webGrouper, which automatically groups each style based on information passed to it from the webTrainer. It can also create new groups if a new style cannot be contained within the existing groups." Millman described webTrainer as the "workhorse" of the vision system. He explained that webTrainer can take any product and automatically train all 51 separate defect detection algorithms to a required sensitivity standard in an unsupervised mode. "It does this by assuming that most of what it sees is good material, but then filters out that which is suspect. To accommodate existing customer quality requirements, webTrainer has optional individual settings per group where it will be 'more' or 'less' sensitivity depending on database settings. This enables webSpector to work as a powerful vision system almost immediately on material it has never seen before." The webCorder enables the entire sheet material to be recorded to disk, every last bit in fact, so that it can be replayed later. This means quality staff can 'skip' to any part of the process very quickly, much like a CD or DVD fast forward function. As Millman said: "It is used to help validate a group, in that results for a newly-created group can be compared with manual inspection results. If a defect is missed by the vision system, the webCorder can be moved to that point and its settings adjusted so that the defect is reliably detected. "The physical material does not need to be re-run through the system, which after a few more goes would probably induce many more defects than there were to begin with. This saves time and scrap and allows good verification of the system's performance," he added. Customers using the system are already seeing benefits. WL Gore and Associates, for example, based in Livingston, Scotland, manufactures high performance PTFE membranes (including 'Goretex') and fabrics for a variety of end uses, ranging from guitar strings and clothing to special uniforms for the MoD. Stuart Speake, associate and (vision) system champion at the company, told Eureka: "The system teaches itself and this has drastically reduced the time to train and validate a new product. We allow it to self-train itself until it has reached a level that is appropriate for our requirements. "We have thousands of different styles, which have to be set up to run. Without the tools provided by the WebSpector self-learning system it would be a mammoth task to introduce new products," he continued. "The software is user-friendly and the fault tracking and statistics generator are useful because they enable us to control quality very tightly. The error mapping allows us to look at and analyse exactly what is happening in our production process. It takes out the subjective-ness and de-skills the process." To date, Shelton has sold three webSpector vision systems. The cost of the system varies from £40k-£50k, up to around £350k for a more sophisticated system. Shelton can even offer clients a 'live support' package where it can inspect a machine over a secure private network. Shelton demonstrated to Eureka the live support of a fibreglass production line at WL Gore's Livingston plant where material was passing the vision system at a rate of 200m/min. The system was installed two years ago and defects were being measured down to 0.5mm What's different about webSpector? The system uses MMX technology in most of its processes. This means there is a lot of processing power available to do a whole array of different image processes which isolate subtle defects that would otherwise be missed and sent to customers. For the general purpose webSpector system, there are nine different defect detection algorithms that can run simultaneously. Six of these use forms of filtering which are notoriously process-hungry operations but which are very effective. The other three processes stabilise the image and cancel out external effects such as variable lighting and factory temperature. This gives better reliability so that the system can inspect to the optimum sensitivity.