Open-Set Recognition of Screen Defects with Negative-Guided Augmented Prototype Generator and Open Feature Generation

In mobile phone screen defect inspection, unknown anomalies may emerge as data accumulates or the industrial environment changes. Open-set recognition (OSR) is needed to classify known defects and identify unknowns. This provides clues to identify familiar problematic processes and warns of the emergence of new problematic processes. However, most OSR methods either lack adaptive and explicit thresholds required for practical applications or are designed for natural datasets with low-resolution images and numerous categories. Such characteristics make them unsuitable for defect datasets. Hence, this article proposes a novel threshold-adaptive open-set classification model incorporating a negative-guided augmented prototype generator. The prototype generator adaptively utilizes diverse information to learn. During inference, it can generate the input-adaptive negative prototype that represents the opposite of known categories. This provides an adaptive threshold with discriminative power against unknowns without calibration. In addition, a feature-focused generative adversarial network (GAN) equipped with a class-wise similarity loss is designed to generate open-set data for high-resolution images. Augmenting the training set with this data enhances the classification model's ability to handle high-resolution defect images. Our method outperforms other OSR methods on the mobile phone screen defect dataset, achieving 99.1% closed-set accuracy (ClosedACC), 99.3% area under the receiver-operator curve (AUROC), 98.6% OSCR, and 96.2% both-set accuracy (BothACC). The code is available at https://github.com/CFZ1/OSR_Screen.