Deep Convolutional Neural Networks (CNN) have shown great success insupervised classification tasks such as character classification or dating.Deep learning methods typically need a lot of annotated training data, which isnot available in many scenarios. In these cases, traditional methods are oftenbetter than or equivalent to deep learning methods. In this paper, we propose asimple, yet effective, way to learn CNN activation features in an unsupervisedmanner. Therefore, we train a deep residual network using surrogate classes.The surrogate classes are created by clustering the training dataset, whereeach cluster index represents one surrogate class. The activations from thepenultimate CNN layer serve as features for subsequent classification tasks. Weevaluate the feature representations on two publicly available datasets. Thefocus lies on the ICDAR17 competition dataset on historical document writeridentification (Historical-WI). We show that the activation features trainedwithout supervision are superior to descriptors of state-of-the-art writeridentification methods. Additionally, we achieve comparable results in the caseof handwriting classification using the ICFHR16 competition dataset onhistorical Latin script types (CLaMM16).
展开▼
