Character segmentation in text line via convolutional neural network

In this paper, we propose a character segmentation method for the task of text recognition. Specifically, our method is designed for splitting the adhesive and incomplete line characters under complex backgrounds. On one hand, we propose to split a text line by training a binary Convolutional Neural Network (CNN) classifier which identifies character images containing a whole character and gap images centering at the middle of two characters. On the other hand, due to the insufficient number of labels in training data, we adopt a weakly supervised training strategy, i.e. P-N Learning, to iteratively update the CNN using the samples that violates the predefined P and N constraints. Experiments on various datasets verifies the effectiveness of our method.