Handwritten character recognition with sequential convolutional neural network

Handwritten character recognition has been an active and challenging research problem, due to the large variations of characters and the dependency relationship between characters. Most of traditional methods rarely consider these two challenges simultaneously. This paper proposes a sequential convolutional neural network(CNN) to solve this problem. It considers handwritten words as image sequences and each node of the sequence is an image of a character. A particularly modified convolutional neural network is firstly employed as a trainable feature extractor to extract low-dimensional, linear separable, translationally invariant features. Furthermore, a sequential model, modified conditional random field(CRF), is utilized to model the dependency between characters. Benefiting from the invariant feature learning ability of CNN and structured prediction ability of CRF, the proposed approach boosts the performance over the standard CRF a large margin. Comparing with the state-of-the-art models, our method achieves a better or comparable performance.