A Comprehensive Dataset for Image Segmentation in Custom Manufacturing Environments

Advancements associated with Industry 4.0 present promising solutions to the persistent challenges encountered by high-mix, low-volume (HMLV) foundries, particularly in process automation. This paper introduces a novel and comprehensive image dataset designed to train an image segmentation neural network for automating the post-processing task of sprue and riser removal for sand-cast parts. The dataset encompasses real camera images, synthetic renderings, and augmented images, alongside a systematic evaluation of image quality metrics and their potential impact on model training performance. By capturing a wide range of part representations and scene complexity, this approach directly addresses the variability inherent in HMLV environments, where conventional standardization is hindered by complex and customized part configurations. The dataset, combined with accompanying quality metrics, provides a valuable resource for evaluating and improving the robustness of machine learning models in industrial imaging contexts. The complete dataset, including turntable control code and image acquisition scripts, is publicly available as an open-source resource on Kaggle.