Under CopyrightKumar, DeepakDeepakKumarHolzbach, GerritGerritHolzbach2024-08-262024-08-262024https://publica.fraunhofer.de/handle/publica/474067https://doi.org/10.24406/publica-359910.24406/publica-3599The assembly of machine parts crucially depends on selecting the correct screws, varying in length, width, and head type (e.g., round, hex). High work rates in manufacturing, driven by cost considerations, often result in manual selection errors. Such inaccuracies necessitate time-consuming corrections, compromising efficiency and elevating costs. This research introduces a novel screw classification algorithm leveraging computer vision and deep learning to mitigate these issues. Central to the research is a deep learning segmentation model for screw identification, combined with custom image processing techniques for classification. This approach utilizes binary segmentation masks to determine screw attributes accurately. The training dataset includes real-world and synthetically generated images, ensuring robust model training. Experimental validation using mono and stereo camera setups demonstrates the algorithm's effectiveness, with stereo achieving a 95.1% accuracy and mono 92.6% across 41 screw types. Additional tests on batch uniformity and incorrect screw identification yielded precision and recall rates of 81.8% and 90%, respectively. These results not only confirm the algorithm's precision and reliability but also its potential in enhancing assembly line productivity, marking a substantial advancement in applying artificial intelligence in industrial automation.enScrew ClassificationScrew SortingComputer VisionDeep LearningAssembly AssistanceObject DetectionIndustrial Automationpaper