More than Deep Learning, Machine Vision Software Advances Across the Board
| By: Winn Hardin, Contributing Editor
As the machine vision industry prepares for its busiest six months of technology events, starting this fall with VISION in Stuttgart, Germany, and continuing into the A3 Business Forum and Automate in early 2019, constituent companies are busily putting the final touches on an exploding number of machine vision software developments. And believe it or not, it’s not all about deep learning. From ultra-high-dynamic-range imaging to self-analyzing algorithms, machine vision customers may come to accept that the secret sauce truly is in the software.
According to Teledyne DALSA’s Development Leader, Jordan Wisniewski, the electronics industry is extremely interested in high-dynamic-range imaging. “As board densities go up and hardware becomes more integrated, electronics customers want to run their lines faster with higher resolution,” says Wisniewski. “As a result, signal-to-noise ratio [SNR] becomes more critical than ever before, especially if you’re trying image layers below the surface through other layers of electronic assemblies.”
The answer is to tackle the problem from both a hardware and software perspective, says Wisniewski. Since Teledyne DALSA makes their own sensors as well as cameras, they can offer customers the opportunity to embed preprocessing algorithms developed either in-house or by the customer that solve SNR challenges for very specific application requirements. Other customers want similar embedded intelligence related to tagging images with metadata that includes preprocessing actions and values to help the host system extract meaningful information from the hundreds of gigabytes of image data produced by advanced machine vision cameras.
“Data reduction is becoming a key concern for many of our medical and electronics customers, and by embedding key algorithms into our cameras for specific customer needs, we’re helping to solve that challenge,” Wisniewski said. “This is almost a new hybrid class of smart cameras that don’t use an integrated development environment but do have some customer-specific processing inside.”
OCR Bulks Up
Datalogic has been known as a leader in image-based machine-readable code readers and one of the few machine vision camera vendors to also offer laser marking systems, but to bring both capabilities into greater alignment, the company has focused on improving its optical character recognition (OCR) and code-reading tools.
“We’re developing tools that can read more challenging codes,” says Bradley Weber, Product Marketing Manager at Datalogic. “Many times characters are not clear black-on-white printing, and many times the print is not perfect or it’s on a varying background.”
Weber adds that Datalogic has also upgraded its pattern-finding tool to be on par with the best in the industry. “We looked at speed, accuracy, and repeatability,” he says. “We’re now the gold standard on all three, able to find objects with the same or better subpixel accuracy as any other player in the market.”
Tordivel AS, maker of Scorpion machine vision software and cameras, is joining Cognex and MVTec GmbH with the introduction of its new NeuralOCR tool that uses deep learning to handle any existing or new font — even when the letters are stamped on metal or raised on rubber and offer no significant visible contrast between lettering and background material. However, unlike other standard neural network-based OCR tools, Scorpion allows the user to draw the font, and then trains the system on the font by creating an image set where various noise elements are introduced into the image. “This way, there is no need for generating all the training images you need for a deep-learning OCR tool,” says Thor Vollset, CEO of Tordivel AS. “We can do both training methods, but we think this method is better.”
When used in conjunction with Scorpion’s new Venom (small- to medium-baseline distances) or Stinger (long baseline) 3D cameras, height information can be added into the 2D image for applications such as stamped metal parts or raised letters on tires.
Meanwhile, Cognex’s deep-learning OCR tool has opened new applications for reading dot-matrix codes in automotive, food packaging, and other challenging applications that were recently beyond the ability of traditional OCR algorithms. According to John Petry, Director of Marketing for Vision Software at Cognex, the new OCR tool ships with a pretrained network that only requires a few customer images to learn a new font.
“No one wants to take the time to label hundreds of images of OCR characters, or deal with the inevitable mistakes,” says Petry. “With our pretrained network, we eliminate those challenges.”
3D and Deep Learning
To improve the ability of its Matrox Imaging Library (MIL) to read 3D text, among other things, Matrox, maker of machine vision cameras and software, has added a photometric 3D algorithm to their toolset. Like Cognex’s SurfaceFX, the Matrox photometric 3D algorithm combines multiple images of the same object using different sources of illumination. By combining the images, MIL can greatly improve the contrast of 3D features, including stamped or raised text.
Matrox has also added a new rectangle shape-finder tool for object location, since most machine vision applications inspect objects based on common geometric shapes. “Simple shapes solve the most common machine vision applications, including squares, rectangles, circles, and ellipses,” says Arnaud Lina, Director of Research and Innovation at Matrox.
Finally, Cognex, MVTec, and Matrox are all introducing image classification tools based on deep learning. Deep learning approaches develop inspection criteria based on the image analysis of images that have been expertly tagged with a description of the defect shown in a given image. This approach is strong for handling hard-to-define defects and product variations, but often requires more time to run and special knowledge to implement.
All three major software companies will be offering new deep learning image classification tools that not only help identify a part or feature in an image, but also provide location information similar to pattern-search algorithms.
As customers have more options for solving machine vision challenges through advanced software, support and training become even more important. For example, MVTec will introduce an improved, easy-to-use HDevEngine for developing scripts in HALCON linking multiple algorithms together. Cognex has integrated its ViDi software with its VisionPro® library, so that customers can easily mix-and-match deep learning with traditional vision tools. Matrox has launched the Matrox Vision Academy online training program to supplement their in-house customer training programs.
Look for more advances on machine vision software at VISION and Automate, coming to a country near you.