Surface inspection is a job performed in many quality controls. It can be for defects
in material which has more than an appearance issue. For rails, dents and marks can have an influence on the lifetime of a rail, and defects can result in catastrophic failure. For structural defects the criteria are often size and position. For appearance it is more complicated.
The first surface inspection job we looked at in 1986 was to check postal stamps. A staff of 20 inspectors looked at each sheet of stamps and inspected in great detail. A small defect could be accepted - but not if it was near the eyes or nose of the queen. The rest was not so important. Today much of the automatic inspection for appearance is done on mass product. Mainly on pharmaceutical devices and their packaging. To receive a medical device with visual defects does not boost the confidence with the product inside. If the manufacturer cannot produce a fault free device, what about the medicine inside? For a very large producer of injection systems, we inspect 500 million devices each year. The systems are installed in big machines where the part is printed and
rotated in front of the cameras. We find smear, spots and inadequate printing.
By measuring the grey levels, we can determine if the correct amount of ink is applied. If the ink is decreasing, we signal the printing mechanism to add more ink. Thereby we have feedback where the printing process is automatically regulated.
A general rule is that spots smaller than 50 mm are acceptable. This limit is derived from the capabilities of a young eye without optical enhancement. If the defects are on a mat and uniform surface it is more or less straight forward. But if the surface is shiny the problems grow. Some of the worst cases is brushed stainless steel. Some medical devices use this nice-looking surface, but to inspect for dents and scratches is difficult. The inspection is mostly done by staff who pick up the part and turn it using the stereo vision humans have. By turning under a light, you can pick up very small defects. This is difficult or at least costly to imitate with a vision system. It will involve mechanics and maybe spotlights from different angles.
Spotlights can be turned of and on in sequence which is a lot simpler than mechanical handling. Illumination of round objects is not easy. A good way to solve this is to rotate the part and use line scan cameras. Thereby, the light will always fall on the cylinder at the same angle, and when the line scan camera has recorded the full circumference you have a completely flat unfolded image of the surface. It looks quite remarkable. To make the vision system analyse the image is then straight forward.
When introducing automatic inspection, we always look at the implications it has for the production. Often inspection is looked upon as a separate function the staff perform. But as vision systems suppliers we always look at what else the workers are doing. It can be fixing jam ups in the mechanics, og cleaning the machinery. When you go to automatic inspection and take staff away from the line, you may miss some important outlying functions. To develop a system, we need samples of good and faulty products. What we want is the smallest defects you want to reject. This is the type of defects you may or may not accept. The borderline samples are the ones that determine the required sensitivity and thereby the cost of the vision system. In many cases the vision system can incorporate Machine Learning which can be taught by feeding many pictures of parts with good and bad surfaces. This learning process can be automated with our annotation software, enabling the operators to do a large part of the teaching. Vision systems for surface inspection can generate statistics and trend graphs.
These are useful to monitor drift and the general quality of the production. You will surely also be able to monitor and compare your suppliers of raw materials, parts or products.Surface inspection systems can be relied upon. They will work for years and do a much better inspection than operators.
You can learn more about, solving a surface inspection task with machine vision, in this case story: