Image Data Augmentation through Generative Adversarial Networks for Waste Sorting

The overwhelming volumes of municipal solid waste (MSW) generated by human activities present a growing challenge for their sustainable management. In materials recovery facilities, where MSW is sorted to recover recyclable materials and remove contaminants, the accumulation of waste often exceeds the capacity of human operators to manage it. Hence, waste sorting greatly benefits in efficiency from the integration of automation into the process. Pixel-wise annotated images of waste streams are essential for an intelligent computer vision system to learn the identification of various materials through semantic segmentation. However, the availability of this type of data is constrained by the labor-intensive nature of manual image labeling. Therefore, proper data augmentation is required to enhance the robustness of the trained segmentation model. Given the recent advancements in generative artificial intelligence, this thesis explores its application in the production of synthetic data for computer vision-driven waste sorting. The groundbreaking framework of Generative Adversarial Networks (GANs), whose widespread development has led to outstanding results in terms of realism of generated images, is employed herein. After successfully training a GAN in the representation of waste streams, a specific method is designed to control its generation process so as to comply with the semantic composition of the input. Semantic segmentation models are then trained on waste sorting datasets synthetically enriched through the developed data augmentation technique; the results highlight the method's effectiveness, alongside some limitations related to the provided variety.