The Design And Implementation Of A Classification And Generation System Based On Recurrent Neural Network

With the rapid development of deep learning, we focus on how to do efficient handwriting recognition using neural network. Inadequate data collection as well as existing data sampling will seriously affect the training effect. If we can take advantage of an excellent learning dataset to automatically generate the trajectory of the character sequence, we can effectively solve these problems. Overlapped handwriting recognition is widely used to input text in smart devices since it allows to write continuous characters on a size-restricted screens. How to segment the stroke sequences into characters is a crucial step before recognit ion.This paper will focus on solving the two problems proposed above. For the first problems that how to expand the data set, firstly we can learn offsets between adjacent points in the good sampling dataset. Then according to the offsets between points in character sequences, we can predict the position of the next point by using the position information of the current point and the offset between current point and previous point. We use mixture Gaussian Distrubition model to describe the distribution relation of the offset. For the second problem, how to segment the stroke sequences into characters is currently formulated as a two-class sequential classification problem merely evaluating on the relationships between a pair of adjacent strokes. In order to deal with this classification problem, each adjacent stroke pair is expressed as a feature vector, which is convenient to train the network. These two problems depend on the long contextual history information from massive data in common, so we have adopted RNN model to train network.The sequence of characters which is generated by us can be recognized using the naked eye and these sequence of characters learn the front styles in trainset. The specificity and precision of the trajectories sequence classification system is 91.04% and 79.86% respectively, which outperforms the traditional methods.