Size-free one-class classification approach using probabilistic local cluster balance

One-class classification (OCC) is a supervised classification problem where the training data is solely one class. However, existing OCC algorithms must resize input images to a fixed size, which may lead to information loss. This paper proposes the first size-free OCC algorithm: one-class probabilistic local cluster balance (OCPLCB). The proposed method can extract feature vectors without resizing. OCPLCB applies k-means clustering to sliding windows in images. Subsequently, a feature vector is extracted by dividing cluster balance by the number of sliding windows. Finally, the feature vectors are classified by traditional OCC algorithms. Additionally, this study avoids hyperparameter tuning by combining multiple hyperparameters. The proposed method shows fair areas under the receiver operating characteristic curve (AUC) scores in grayscale datasets without deep learning. The discussion section provides the benefits of size-free OCC, limitations, and potential future work. Source code is uploaded at https://github.com/ToshiHayashi/OCPLCB