Anomaly Detection

Defect classification needs many examples of every defect type, yet faulty products are rare on a production line and a new kind of defect can appear any day. In anomaly detection the model learns what a good product looks like and gives every new image an anomaly score; a product scoring above the threshold is rejected as NOK, and the location of the defect is shown with a heat map.

We train the model for each application with images of good products collected on site and develop the software tailor-made for the customer's line. Defects such as dented or torn packaging, missing labels, stains or wrong placement are shown to the operator visually, recorded, and the faulty product is sent to the reject line.

One of our applications on the shop floor is dent and tear inspection on sugar boxes, and we have also carried out anomaly detection on brake pads. The image on this page is a demonstration prepared from our sugar box work: on the left a good box is rated "good" with a low score, on the right a box with a defect at its bottom is flagged as an "anomaly" with a high score.