Neural network based over-segmentation for scene text recognition

Over-segmentation is often used in text recognition to generate candidate characters. In this paper, we propose a neural network-based over-segmentation method for cropped scene text recognition. On binarized text line image, a segmentation window slides over each connected component, and a neural network is used to classify whether the window locates a segmentation point or not. We evaluate several feature representations for window classification and combine sliding window-based segmentation with shape-based splitting. Experimental results on two benchmark datasets demonstrate the superiority and effectiveness of our method in respect of segmentation point detection and word recognition.