Read complex and challenging codes under any condition with deep learning-based OCR
Imagine the time savings of a pre-trained optical character recognition (OCR) and verification (OCV) tool. By eliminating arduous training upfront, this tool would work out-of-the-box. And if an unrecognizable font did appear, likely because of specular glare, poor etching, or another cause of deformity, that tool could be retrained on the factory floor, and only on the problematic misread characters. That’s the promise of deep learning-based OCR and OCV tools, which rely on a pre-trained, omnifont library and only require training on application-specific fonts.
Consistent, easily readable fonts set against a light or dark background can be solved well enough with traditional machine vision. But manufacturers need a new kind of dedicated OCR solution for when confusing backgrounds and imaging issues challenge their machine vision systems, and when the number and type of application-specific fonts are unpredictable. Badly deformed, skewed, and poorly etched codes are some of the common culprits. Other tricky codes are those involving low-contrast characters or codes printed, etched, dot-peened, or embossed on confusing backgrounds.
Let’s explore how deep learning-based software offers a radically easy, accurate alternative to these complex OCR/OCV inspection applications.
Deep Learning OCR for the Automotive Industry
Automotive manufacturers as well as parts suppliers use serial numbers to track high-value parts through the supply chain and ensure they are matched with the correct assemblies. Many parts contain ten-digit, dot-peened serial numbers so that when errors occur during testing, the source can be traced.
In the case of a recall, affected parts can be quickly taken off the market. In addition to typical readability challenges—parts of the serial numbers may become abraded during the casting or sanding processes, for example—the glaring metallic surfaces can sometimes confuse the camera in an automated inspection system. When dot-peened serial numbers become grossly deformed and illegible, this slows down the OCR and OCV processes and threatens effective traceability.
Deep learning-based OCR/OCV tools are well-matched to these challenges because they rely on pre-trained, omnifont libraries to identify even the most difficult-to-read codes out-of-the-box. Cognex Deep Learning only requires application-specific adjustments, meaning that instead of training an algorithm to identify each code’s individual character or number, a training engineer—not a vision expert—simply defines the region of interest (normally the area with the misread characters), sets the character size, and labels the image. Any misread characters or application-specific fonts can be easily retrained on the factory floor.
Deep Learning OCR for the Electronics Industry
Laser-etched codes on electronic components like integrated circuit (IC) packages and lead frames are essential features for all electronic hardware manufacturers. These barcodes and serial numbers contain information about when and where parts were manufactured, their lot numbers, and testing data. They may also encode information about solder temperature and flux density—essential information as components are mounted onto chips and assembled into modules
These codes are read at every stage of value-add up through final assembly and device testing to ensure that hardware is assembled correctly and contains the right components. Given the small size of most semiconductors and the space constraints on a PCB board, a manufacturers’ identification system must be quite robust to keep production running at full speed and keep track of high-value components. This is also true for finished device manufacturers, who routinely have to read laser-engraved codes on the sides of slider heads which can be as small as 1.1 mm x 1.4 mm. Unsurprisingly, laser-marked codes can degrade during production and become difficult to read.
In these situations, deep learning-based OCR/OCV technology offers an out-of-the-box solution, skipping over tedious hours of training thanks to a pre-trained, omnifont library which recognizes even deformed and skewed characters. An engineer can quickly make application-specific adjustments and re-train on misread codes. Productivity benefits are immediate as no-reads decrease and machine uptime is maximized.
Deep Learning OCR for the Packaging Industry
Manufacturers must have reliable systems in place to recognize and verify the trail of information which follows each packaged good through the supply chain. In their efforts to ensure full package traceability, food and beverage and consumer products manufacturers sometimes confront challenging codes.
The usual suspects are either low-contrast characters printed on label-based packaging codes or deformed, embossed characters on injection-molded parts, such as a bottle cap. The codes may be used to match multi-part packaging or, more likely, contain date/lot codes that embed information about contents, origin, and date of manufacture. In these cases, manufacturers rely on OCR/OCV equipment to quickly locate affected products and pull them out of production or off shelves. Traditional OCR/OCV technology requires upfront training to learn various fonts and even then can struggle to decode poorly contrasted characters.
Deep learning-based OCR/OCV technology doesn’t have the same limitations and is able to read most poorly contrasted letters and numbers automatically. By embracing deep learning-based technology, manufacturers can keep abreast of food safety and traceability laws and facilitate recalls with minimal impact on their production.
Deep Learning OCR for the Life Sciences Industry
Effective OCR and OCV is critical for the life sciences, which has tightly regulated track-and-trace laws. To remain in compliance, manufacturers as well as hospitals need to scan the codes on medical devices, surgical tools, and patients’ hospital ID bracelets at every point of handling and point of use. This allows them to keep a tight hand on devices and pharmaceuticals throughout the supply chain in case of a safety event.
With so much equipment and many people touching products, codes can become deformed and skewed. Imaging quality may also be subpar, changing the code’s appearance to the camera. Rather than investing time to train a machine vision system to recognize the panoply of fonts it’s likely to encounter, the life sciences industry is turning to deep learning-based image analysis software to do the work for them.
Deep learning-based OCR tools are effective and easy to deploy on codes printed onto challenging substrates and prone to deformity like dot-peened codes on metal parts, embossed characters on injection-molded products, label-based codes on packaging, and laser-etched codes on electronic components. Cognex Deep Learning's OCR/OCV technology recognizes most alphanumeric text out-of-the-box and only requires brief upfront training to set the region of interest and character size. The system can retrain on misread characters quickly on the factory floor, so that manufacturers don’t miss a beat.
To find out how deep learning is revolutionizing factory automation applications like OCR and OCV, download the free whitepaper, Deep Learning for Factory Automation: