Cosmetic Defect Inspection

The high-end cosmetics produced by S customer are very expensive, and even the bottles need to be presented perfectly to customers. The difficulty of cosmetic bottle inspection lies in that the transparent bottle body needs the assistance of specific lights to see defects, the production line lacks defect data, and the shape and color of the defects are uncertain.

In response to the above difficulties, Aternas helped customers design an auxiliary lighting device that can accurately capture images of defects through cameras. In order to solve the problem of lack of data, Aternas designed a virtual data expansion tool based on the pattern of existing data, adding a variety of defect data for training neural networks. At the same time, in order to make the model easily deployable at any location on the production line, Aternas designed a lightweight neural network and performed a distillation operation on the trained model, so that small models deployed on the end side can also have excellent performance.