Inspection of automotive exterior parts

We have often been asked to inspect automobile exterior parts for some time.

However,

• This is a curved and multi-surface inspection, which makes the setup quite cumbersome.
• Lighting conditions under which defects are visible are almost exclusively limited to the "side of the lighting reflection".
• High gloss. Also pear-shaped with metallic coating.

The level of difficulty was extremely high, and we could not compete at all.

We have found a way to visualize defects by devising lighting, but the images are quite unstable and we have not found a way to determine if they are good or bad. (We will refrain from publishing the images on our blog because we have the know-how here.)

In such a situation, I tried with a faint hope that deep learning might be able to extract only defects from unstable images.

• THE POSITIONAL RELATIONSHIP BETWEEN THE CAMERA AND THE OBJECT VARIES. (IF ALL PARTS CAN BE INSPECTED BY A COMMON AI ENGINE, IT IS POSSIBLE TO HANDLE MULTIPLE TYPES AND SURFACES.)
• Annotation (labeling) of defective parts of defective images. less than 100 sheets, about 1 hour's work.
• Targets foreign objects and uneven paint (with slightly uneven surfaces).

I think it's impossible. I don't think it's possible." I don't think it's possible.

But...

• ALL DEFECTS IN NG IMAGES ARE DETECTED. (THIS INCLUDES IMAGES USED FOR TRAINING.)
• IN THE OK IMAGES, ABOUT 70% OF ALL PIXELS ARE OK. THERE WERE A FEW SMALL OVER-DETECTED PIXELS WHOSE REASONS WERE NOT CLEAR, BUT THERE WERE ALSO "UNEXPECTED DEFECTS" AT THE LEVEL OF "DO YOU EVEN FIND SUCH THINGS? HOWEVER, THERE ARE ALSO "UNEXPECTED DEFECTS" AT THE LEVEL OF "ARE YOU GOING TO FIND SUCH A THING? WE EXPECT THAT OVER-DETECTION CAN BE SUPPRESSED IF THRESHOLDS FOR "CONFIDENCE LEVEL" AND "AREA" ARE IMPLEMENTED.

The result was "No way. No way."

Originally, we wanted to unveil it at an exhibition, but the exhibition was cancelled. We consider it positive that we are glad that we did not have to open up our expertise in defect visualization.

If you are interested, please let us know and we will send you "raw images" and "detection results".