Semi-supervised Sorting via Deep Feature Extraction and Density Based Clustering with User Feedback

Efficient data sorting remains a significant challenge when faced with unknown object characteristics or appearances. Unsupervised learning struggles with contextual nuances and subjective human perspectives, complicating the sorting process. In this paper, we propose a novel semi-supervised algorithm merging deep feature extraction, density-based clustering, and user feedback to address this complexity. Our approach utilizes a deep feature extractor, followed by dimensionality reduction techniques for refined features. User feedback, collected through intuitive queries, aids clustering by applying cluster splitting, merging, and outlier assignment. The resulting labels are then used to train a multi-class support vector machine in the original feature space. We demonstrate how little user feedback reduces classification errors on unseen data by up to \(75\,\%\) raising classification accuracy to nearly \(90\,\%\) . We demonstrate the algorithm’s potential through a screw sorting demonstrator, showcasing its adaptability and efficacy in real-world scenarios. The proposed system, with its intuitive user interface, minimizes interactions while advancing knowledge. The integration of user feedback ensures precise object sorting, emphasizing the algorithm’s versatility across diverse image datasets and real-world applications.