Office Documents Classification under Limited Sample

A Case of Table Detection Inside Court Files

Abstract

 Deep convolutional neural networks (CNNs) became an industry standard in image processing. However, in order to keep their high efficiency, a large annotated sample is required in the case of supervised learning. In this paper we apply the techniques specific for relatively small sample to a court files dataset. Specifically, we propose transfer learning and semisupervised learning to classify scanned page as having a table or not. We use four CNNs architectures established in the literature and find that transfer learning improves the classification performance, compared to the fully supervised learning. This result is especially evident in the scenarios where only a part of convolutioanl layers are transferred. The gains from semisupervised learning are ambiguous, as the results vary over CNNs architectures. Overall, our results show that office documents classification can achieve high accuracy when transferring initial convolutional layers is applied.