A human eyeball and a camera, each placed over a torso with fists raised against each other

New Employee or AI?

This is the scenario: you need to add inspection for scratches to your production line. You have the choice of hiring a worker or using machine vision. Which option works better?

New Employee
Machine Vision/DL

You hire a new employee.  After the onboarding process, you take them to the production line and assign them to inspect each product for scratches and reject the ones that are scratched.

How can this work?

The new employee brings with them a pair of eyes plus a wealth of knowledge from life’s experiences.  They have seen all kinds of scratches from deep gouges in furniture to barely noticeable scratches in a family heirloom.

With machine vision, it’s not so easy. Even using deep learning (DL) along with transfer learning where the DL model is pretrained with a similar task, there is still a need to collect and label sample images and train for the specific task.

Of course, DL doesn’t come with a camera and one that’s appropriate must be supplied and mounted.

At the end of the first shift, you notice that the quantity of reject parts is much greater than expected.

Part of this may be because the extent of the problem was not known. Another cause is poor training of the new employee as to what was and was not a rejectable scratch.

In a sense, the employee is relying on transfer learning just like DL.

At the end of the first day, you notice the machine vision system is functioning, but not as well as expected given the false accepts and false rejects.

More training is required. Also, the imaging – camera and lens resolution and lighting – may not be compatible with the application.

Also, at the end of the second day, you notice that the new employee is limiting production due to the time they take to inspect the products.

Watching the working, you observe that they pick up each product and wobble it around to make scratches more evident. This is a normal technique employed by people when performing inspections.

Without wobble, it’s unlikely the worker can perform inspection with the performance needed.
Also, the time used to wobble the part for inspection momentarily stops the production line or necessitates it to go more slowly.

While the capability to wobble a part while viewing it is not currently available in automation, photogrammetric stereo is available and gives much the same result as the human wobble.

Photogrammetric stereo requires the production line to be stopped while four successive images are taken, and the imaging can be accomplished in about one-tenth of a second.

With precision motion and strobed light sources, tracking photometric stereo can be performed with the part moving. This adds significant complexity to the solution.

The new employee may have young eyes that can see without glasses or may have older eyes that need glasses and may also have cataracts that limits visual acuity (resolution).

Some finer scratches may be missed with older eyes that would be found with young eyes.

Human eyes only have maximum resolution in the fovea, a very small area of the retina. This causes inspection for fine details, such as scratches, to be slow.

Image resolution is important in imaging with DL just as it is with conventional machine vision. If the scratch is not resolved in the image, it will not be found by image processing of any method.

Some people are under the impression that DL compensates for inadequate imaging. This is shown not to be true.

The machine vision camera together with a lens matched to it, has uniform resolution across the entire image. There is no speed penalty in finding very small defects such as scratches.

The new employee complains that the illumination in the work area is bad. It is somewhat dim creating eye strain and making seeing fine scratches hard. And there are a few bright lights that shine into the worker’s eyes making eye strain a big problem.

The solution is to redesign the illumination in the work area for good inspection. Light should be reasonably bright, with no light glaring directly or indirectly off the product into the worker’s eyes.

Machine vision requires good lighting for adequate exposure to minimize image noise and to minimize glare that creates areas of no information in the image.

While DL will attempt to compensate for poor illumination with more extensive training, it will never perform up to its potential when used with well-designed lighting.