I have a very good friend; let’s call him “Jake.” Jake has spent his entire adult life in manufacturing and supply chain. When his employer decided to develop a customer-centric marketing strategy, they called several large global firms known for their marketing strategy development. Each firm presented their impressive, winning sales presentations highlighting their proprietary methodologies, focused organic growth models, customer profile development, and new business model strategies. Of course, each of these polished presentations arrived with a 7-figure price tag to deliver the desired services. Jake, no stranger to sarcasm, said “Hey, I have an idea. How about we ask our customers what they want, and then do that?” Sage advice. Jake’s no-nonsense approach “won” and that is exactly what his employer did . . . with very successful results.

I always keep this story in mind when visiting customers and prospects. It is enlightening the things we learn from people who spend their working lives thriving in the constant whirlwind of the manufacturing world. Over the past year, I have had the opportunity to meet with the following companies:

  • Global glass manufacturer with an interesting problem to solve in their automotive glass division. Their existing defect detection system created several challenges. It could not reliably differentiate between naturally occurring dirt, grime and water spots from real defects that are real defects. In other words, the system generated an excessive number of false positives. For this reason, they could not equip their production lines with pick and place robots. They also need a large number of trained operators to override the vision systems erroneous decisions.
  • Automotive interiors manufacturer purchased a sophisticated vision system that could not differentiate surface anomalies from real defects. For example, the existing vision system confused lint sitting on top of the fabric for yarn pulls. With the aid of additional human inspection for false positives, their vision system would reduce yield and, consequently, reduce revenue. For this reason, they had human inspectors oversee, and override decisions made by the system. Lots of them.
  • Global tire manufacturer has a visual inspection station at the end of the tire building process. The vision system captures images of the exterior and interior of the tire. Because of the nature of tires (black on black), their false positive and false negative rates are unacceptably high. Consequently, there are human inspectors who must look and feel the tire by running their hands over every square inch of the tire, inside and out, feeling for defects.
  • Pharmaceutical company with a chemical process discovered that when filling vials, occasionally foam would form in the vial. The vision system misinterpreted foam as particulate matter in the liquid or glass defects, both causing an unacceptable level of false positives. Additional human effort must be expended separating the false positives from the true.

Listening to these customers, I discovered a common theme. For the past two or three decades, manufacturers have made significant investments in machine vision systems to automate defect detection and classification. In each example, we found that the cameras, optics, lighting, and image capture was of sufficient quality. Unfortunately, the software scoring the images was not. Modern deep learning and neural network technology can dramatically improve the quality of the scoring results. For example, with one customer we were able to demonstrate a reduction of false positives from 29% to less than 1%. The customer’s existing vision system was highly specialized and represented a significant investment. Replacing the vision system was not an acceptable solution. Our solution to the problem was to create a deep learning model that scored much more accurately. We operationalized the model by creating a system to intercept images from the vision system. We then scored the image using the improved model. Once scored, the results are shared on the operator’s console. This is a preliminary step, however, as the longer term goal is to install pick and place robots. Once that is accomplished, our solution will communicate pass/fail commands directly to a robot.

Our customers are happy to learn we can extend the life of the investments made in their existing vision systems and at the same time, reduce the amount of manual intervention required to compensate for image scoring software.

Do you have an existing vision system with high degree of false positives? Do you know the frequency of defects by defect classification? Do you need a cadre of people to monitor the results of your inspection system?

If so, we should talk. Contact us.


Phil Morris | CEO & Co-Founder

Phil is responsible for defining and executing the business strategy and strategic alliances for Mariner. Phil has spent 25 years in manufacturing information technology responsible for both administrative and factory floor systems.