Deep Learning

Classical machine vision requires describing every defect type by its features one by one, an approach that quickly reaches its limits on variable textures and patterns. In deep learning, many examples whose class is known are given to the system, and a convolutional neural network (CNN) works out what they have in common and what sets them apart.

We train the models on each customer's own product and defect images. Training runs on our own GPU machine; for customers whose product and defect variety will grow over time, we also install a GPU computer with a training program. Newly trained model weights are loaded into the software we installed and the system adapts itself to the new situation. Once trained, the model decides on every image within milliseconds on the line.

We use deep learning in many of our projects running on the shop floor: quality inspection of automotive headrests, telling broken glassware from intact pieces on a belt, reading DOT codes on tire sidewalls, identifying the type of mixed products in 3D images and segmenting panel radiators on a conveyor. Fabric defects (oil marks, weft/warp misalignment, tears), flaws in the natural texture of leather and chips on glass bottle rims are also among the applications we have solved, and we are developing a deep-learning-based inspection machine that finds every type of fabric defect.