Potential of Deep Learning methods for image processing in sensor-based sorting: data generation, training strategies and model architectures

The main component of a sensor-based sorting system is an imaging sensor and the associated data processing unit for detecting and classifying bulk material objects. High occupancy densities and objects with similar appearance lead to increasing problems for conventional image processing algorithms in object and class separation. Therefore, in this article, specialized Deep Learning approaches were applied to two datasets for instance segmentation. Due to the need for a large amount of training data for such models, a method for synthetic training data generation has been developed. Subsequently, established model architectures as well as an own approach specialized for the problem characteristics is presented and compared regarding their detection performance. Finally, the models are evaluated in terms of their speed and therefore their potential use in a sorting system. Our approach more than halves the inference time of the fastest model while achieving the best detection performance.