Detecting Kuzushiji Characters from Historical Documents by Two-Dimensional Context Box Proposal Network

Abstract

Detecting characters from historical documents is a challenging problem due to the cursive and connected characters. It is a fundamental task for recognizing historical documents. In this paper, we propose a two-dimensional context box proposal network. The network has three parts: feature extraction, two-dimensional context, and box proposal regression. For feature extraction, we employ VGG16 to extract features from an input image. Then, the extracted features are processed by two-dimensional context. We employ two Bidirectional Long Short Term Memory (BLSTM) to explore vertical and horizontal meaningful context. Finally, the bounding box is predicted from the output of the two-dimensional context. We tested our proposed system on 28 books of Kuzushiji documents. The results of the experiments show the effectiveness of our proposed two-dimensional context for character detection. Our system achieved 88.50% and 92.46% of F1 score on validation and testing sets, respectively, which outperforms the baseline system (Connectionist Text Proposal Network).