Handwritten Characters Recognition Method Based On Deep Learning

Recently. deep learning has been the frontier of machine learning and artificial research.After 2012 a large number of state-of-the-art results on recognition have been reported, and almost all of them were based on deep learning method. In this thesis, I mainly focus on deep learning and its application on handwritten characters recognition, which is not a new topic,but it is very challenging. Now researchers have proposed many algorithms that can recognize single handwritten character, and their performance on MNIST database was satisfying, some of them even perform better than human. However, it is still unsolved in terms of recognizing a string of handwritten characters. In this dissertation, I endeavor to apply deep learning method to solve this problem.Firstly, I develop an energy model which is able to restore occluded images. At the test stage. this model performs very well. However, when it is used to restore big images, the perfor-mance is unsatisfying. Therefor, I apply convolutional method to improve its performance. In the convolutional version, the model only uses local features instead of the entire image, which is more reasonable, because it is believed that the value of a pixel is highly related to the nearby pixels.Secondly, I research on recognizing single handwritten character and compare several algorithms. Of cause, deep neural networks perform the best and unsupervised pre-training is relatively powerful. Thereby, I apply unsupervised pre-training method to train convolutional neural networks, developing two unsupervised algorithm. After pre-training, we gain better results and the test error sees a considerable decrease.Thirdly, in this dissertation, a framework based on over-segment and deep learning for recognizing a string of handwritten characters is proposed. To segment the images, I develop a simple but efficient algorithm that is capable of splitting raw images into small parts. In this algorithm, I used low-pass filter to reduce the number of small parts and complexity. After over-segment, deep learning method is used to evaluate each part, generating an interpretation graph.At next stage, we use Viterbi algorithm to find the best path with low penalty and this path is associated with labels that we want. Finally, we come up with a global training algorithm to train the entire system.Last but not least, in this dissertation, I also conduct many experiments on deep learning which is essential for the following studies. We can easily find that deep learning is able to learning more significant features than other methods and the representation is very helpful for classification and recognition.To sum up, this dissertation address the problem of handwriting recognition and develop a hybrid model to solve the problem.