Synthetic data generation based on grid deformation for waste recycling applications
Nick Tsagarakis, Alexandros Antonaras, Michail Maniadakis · 2023
Today, waste recycling is supported by intelligent robots that use machine learning to identify and sort recyclables. The development of computer vision applications based on machine learning relies heavily on large datasets that are used to train deep neural network models. In recent years, methods that allow the creation of large training datasets from a limited initial set of images have been investigated. This paper describes a method in which segmented images of real recyclables (polyethylene terephthalate, PETE) are artificially deformed using mesh transformation to create new instances of the recyclable objects. The new instances are placed on real backgrounds to create synthetic images. This process allows the generation of large artificial datasets used for training neural networks. We evaluate the usability of these datasets by studying the extent to which they can improve the performance of trained models when applied in real and challenging industrial images. In particular, we consider the main metrics used to evaluate the performance of classification models, namely Accuracy, Precision and Recall. The results obtained show that including even small-scale object deformations in the artificial datasets can slightly improve the Accuracy and significantly improve the model Recall, while Precision of the model remains unchanged.