A deep learning based character recognition system from multimedia document

摘要

Text recognition from natural scene images is very tough task now these days compare than videos. Application of image processing called pattern recognition make easy to recognize text from multimedia documents. A pattern can be fingerprint image, handwritten word sample, human face images, speech signal and DNA sequence etc or we can say that all pattern are in machine editable form. Text can be recognized with and without segmentation of character. Segmentation can be line, word or character level and without segmentation character is recognized from whole text image. Character recognition is a field of research and various research has been done in the area of pattern recognition. There we use a new technique called diagonal based feature extraction in last layer of convolutional neural network and make feature extraction easy with the help of genetic algorithm. After extraction of feature we provide training to extreme learning machine. Along this feature extraction technique we use feed forward network as a classifier and convolution neural network for feature extractor. It is a deep learning based technique of neural network which use for classification or recognition of text. This is basically used for providing training and in testing phase. CRConvNet has more layers working of all layer shown in flowchart. One dataset which contain 360 training set data that are all in capital(A-Z) and small(a-z) alphabet, digit (0-9) and some special character are also used. Another dataset contains samples of video and images (ICDAR 2003) for testing. Extensive studies shows that the recognition system which using diagonal based feature learning provide high recognition accuracy while requiring less time for training.