AI Learning at the Edge — Application to Integration-Free Quality Control

EMVA Business Conference 2020.
Virtual Conference including Online-Networking and the pivotal event for the machine vision and image processing industry.

Today the mainstream of Artificial Intelligence (AI) is based on fully supervised deep neural networks. Mainly used on classification tasks, they brought a major disruption in terms of performance with respect to traditional rules-based computer vision solutions. Nevertheless, because of the complexity of data gathering and supervision, they still require specialists, i.e. data scientists to train the solution offline and only inference is applied on the edge. As a result, most of current AI solutions fail to deliver the promise of artificial intelligence, i.e. a system that is able to be trained and operated directly by the end user. While this may still be a good answer to issues involving generic data like cars or pedestrians for which huge datasets exist on the cloud, we believe that applications where data is very scarce and specific to the customer require a new approach, a third generation of Artificial Intelligence. One significant example is the automation of quality control of manufactured goods. The parts manufactured by a factory are usually very specific to this given factory and usually very scarcely available at the beginning of the production cycle. Defective parts are (hopefully) a rare event, that is very difficult to gather and supervise in sufficient quantities.
To solve this issue, AnotherBrain designed a product that can be operated from day one directly by the factory with no computer vision or data science knowledge. Supervision of the data is very limited and mostly automated. Training is continuous and very simple to be performed. To this end, we created a solution able to extract all the pixels of the part from the background and spot them for anomaly using only good parts as a training set. Furthermore, we found a way to select only a few pictures among the production data that is used for continuous improvement of the system. Finally, we propose an original way to localize and segment defects without supervision.

This talk is dedicated to the review of this solution and its technical contributions to the field of automatic quality control.


The EMVA Business Conference unites business leaders and technical experts within the machine vision industry from Europe and around the world to present and discuss the trends and challenges of our business. It provides a unique platform in Europe for networking, establishing contacts and exchanging ideas. The EMVA Business Conference sets the course for developing and strengthening our industry in Europe.

Join CEOs, managing directors, corporate strategists, marketing directors, technical managers and other executives of our business to exchange market intelligence and news of innovative technologies.