On the generation and assessment of synthetic waste images

In contemporary waste recycling, the assistance of autonomous robotic systems, equipped with machine learning capabilities, has become crucial for the identification and sorting of recyclable materials. The evolution of computer vision applications, reliant on machine learning, heavily hinges on extensive datasets employed for training intricate deep neural network models. Recently several works from various fields have explored techniques that facilitate the generation of big synthetic datasets starting from an initially limited set of images. This paper proposes a novel approach for generating synthetic waste images, which involves two main steps. The first regards the use of a neural network to implement a "critic" that can evaluate how realistic, synthetic images of recyclable objects may be. The second involves applying multiple random elastic deformations to images of individual recyclable objects to generate a large number of new appearances of the given objects. The critic evaluates the generated images, gauging their realism through a confidence score. We employ a rigorous confidence threshold to identify synthetic images with a notably realistic appearance. These individual object images are then utilized to craft composite synthetic images depicting multiple recyclable objects on a conveyor belt transporting recyclable waste in an industrial setting. The above summarized process facilitates the creation of expansive artificial datasets crucial for training neural networks. The efficacy of these datasets is assessed by examining their impact on the performance of trained detection models when applied to previously unseen and challenging industrial images. The obtained results show that the use of the synthetic datasets leads to better classification models in terms of both precision and accuracy, motivating more research in the field of artificially generated datasets.