Semantic Segmentation

Classical thresholding and edge detection become unstable on glossy surfaces, under changing light and against busy backgrounds. A deep-learning semantic segmentation model learns to recognize the product from labeled examples and extracts the area the product covers in every image as a mask.

We train the model for each application with images collected and labeled on site; training runs on our own GPU infrastructure, and the trained model is loaded into the software we installed to run on the line. The outline and position found are used by the robot to pick the product.

An example from the field: on a panel radiator line, this is how we find the radiator the robot will pick. Even on the glossy, striped surface of the roller conveyor the model separates the radiator from the background, reaching over 99% overlap (IoU) for the radiator class on labeled test images.