Packaging AI-based Inspection System
SEA Vision Group has developed a new AI-based lipstick inspection system. Through the automation of processes still largely carried out by humans, this solution aims to improve the packaging of a product currently considered one of the world’s bestselling cosmetics.
Ever since its foundation, SEA Vision Group has concentrated its efforts in the field of industrial vision, aiming to improve continually and to overcome the limitations that typically emerge in packaging and automation processes. Lipsticks — some of the most versatile, eclectic products, with complex quality control — pose very specific problems that require innovative solutions to enable their automated inspection.
Lipsticks come in countless colours, finishes, shapes, formulations and combinations, and many of them, especially in the luxury sector, feature logos and sophisticated decorations impressed on the body or tip of the product.
These distinctive styling features — together with the complexity of the industrial process required to perfectly amalgamate the pigments, oils, waxes and emollients in its formula — mean that lipstick is complicated to produce and to verify in terms of product quality. Even today, in spite of all the efforts made during production to prevent and control a whole series of potential defects, some flaws still pass through quality control undetected. This leads to costly reprocessing: in the worst-case scenario, these products make it onto the market, implying serious risks for businesses in terms of brand reputation.
Artificial Intelligence sets out to remedy a large proportion of these potential defects. The SEA Vision Group system (under development by a joint team from SEA Vision Group and ARGO Vision) uses the semantic segmentation of the areas of the lipstick (e.g. body, tip, neck, mechanism, etc.) to identify every possible flaw pixel by pixel. This is achieved by classifying areas by categories, each of which is assigned a name or “label”. Each part or area of the image is classified by categories and identified by a colour on the screen to provide the operator with immediate information about the areas being inspected.
The system self-learns how to discern an ever-increasing variety of more and more complex defects, item-by-item. Self-learning takes place both on the basis of proprietary datasets — a mix of real and synthetic images generated with the most advanced data augmentation and neural generation techniques - and by combining the different models and parameters learnt over time.
These deep learning-based semantic segmentation techniques, now the de facto standard in the Artificial Intelligence field, greatly accelerate the development of capabilities for the analysis of objects. In other words, the ever-expanding range of scenarios the system can consider enables the system itself to evolve and become more and more precise.
The learning process defines the quality control algorithms and continually evolves to generate new versions of constantly increasing sophistication and autonomy. In this specific issue, the final goal is to generalise the “concept” of lipstick, making the algorithms more and more specialised in quality control regardless of the product’s possible shape, colour and texture.