Sheet metal localization using deep learning and synthetic data

Improving the accuracy of sheet metal localization in industrial machines is of great interest to many automated manufacturing systems. Current vision-based systems typically rely on traditional image processing algorithms to locate the position of sheets in images. However, these algorithms often do not generalize robustly in real production setups. To achieve this, we propose a novel framework consisting of two deep learning models that locate sheets based on their corners, and a data generation pipeline capable of creating the annotated data required to train the models. Evaluation of this framework on real production data shows that the proposed approach locates sheet metal corners highly accurate with an average error of 2.17 pixels, which is at the edge of the theoretically achievable limit defined by the human annotation error in the test dataset. Extensive experiments show that the proposed framework generalizes well and can therefore be used as a backbone for various automated systems for which sheet metal localization is a relevant task.