Machine Vision for Aesthetic Quality Control of Reflective Surfaces

Systems for automatic inspection of product quality are in high demand. However, their prevalence are limited by the expense and difficulty involved in their development. Since inspection systems must be engineered to specific products and environments such systems are generally only viable with long running/high volume product series. The expense and difficulty can be limited by carefully controlling the variation that can be encountered. If the variation can be limited, simple handcrafted algorithms may be enough, otherwise machine learning provides the more expensive and capable alternative.
Inspired by human visual inspection of highly reflective brushed aluminium objects we capture sequences of images across multiple view points. This enables us to employ a spatio-temporal weighting of defects, where defects that occur consistently across view points are considered more severe. This is compared to the confidence scores produced by our defectdetector (YOLOv5). Both methods for weighting the severity of defects are then used to classify objects as either bad or ok. Our results show the challenges with training object detectors on a severely unbalanced and ambiguous dataset. Despite the poor detection performance and difficulty in distinguishing between desiredpatterns and defects, our method proves to classify our small test set with an area under the precision-recall curve of 0.665.