SSS-Net: A shadowed-sets-based semi-supervised sample selection network for classification on noise labeled images

Sample selection is a fundamental technique utilized in image classification with noisy labels. A plethora of sample selection approaches published in the literature are based on a small-loss strategy, in which division thresholds are set manually and the correlation between sample losses is ignored. Furthermore, one of the most evident shortcomings of these approaches is that noisy samples with low-quality pseudo-labels can negatively impact the model resulting in poor performance. In this study, a shadowed-sets-based semi-supervised sample selection network called SSS-Net is developed to address these limitations. Our approach leverages a novel technique that combines a loss-similarity-based-clustering method (LSCM) with the shadowed-sets theory to adaptively select clean samples. We then introduce an original high-quality pseudo-label sample reselection (HPSR) strategy, which is designed through the co-training of two networks, to pick the samples with high-quality pseudo-labels. Finally, the selected samples are utilized to further train the network and complete classification. This study presents an automated approach that determines optimal division thresholds to select clean samples adaptively. Furthermore, it improves the current semi-supervised sample selection method by effectively utilizing noisy samples. The suitability and promising performance of the proposed approach are supported through experimental studies using five real-world datasets. Comparative studies involving several state-of-the-art methods are also reported.