Text Recognition In Natural Scenes Based On Convolutional Neural Network

Text in natural scenes contains more information, people can understand the image by using these information.This paper focuses on the problem of text recognition in natural images. The problem is more challenging than reading text in scanned documents, and the word recognition in natural scenes has more practical value. Based on previous work, this paper proposes a natural scene text recognition system based on deep learning.The natural scene word text recognition system contains four parts.The first part is the massive sample generation module, the second part is the data and processing module, the third part is the text detection module, the fourth part is the text recognition module.In the sample generation module, this paper analyzes the characteristics of the text in natural scenes, such as the impact of light, font deformation, shadow, blur, noise and so on.Then we use these features to make a text generator, in the process of making the text generator, we collected the approximately 30000 picture, and remove these pictures with text.We also collected230 different fonts, five thousand common words. We generated three million character training samples, which will make it possible to train the network.In the image preprocessing module, there are three processing steps. The first is the contrast normalization, this step is mainly to ensure that the image of the image data in the vicinity of zero.Image preprocessing of the second step is the principal component analysis method to reduce the dimension, and the third is ZCA whitening, the main objective is to reduce the correlation between the two dimensions of the image pixels.In the detection module of this paper. Firstly, we trained a two class classifier with five volumes of multilayer neural network. This classifier can distinguish characters and non characters of image block.The post processing part of the text detection is calculating the position of the text box containing the text according to the response points of the character blocks.The recognition module module also uses a convolutional neural network, because the character classifier has 62 categories, so this paper in the text recognition module use the CNN with the 7 volume of the multilayer neural network. In the postprocessing section, we use the BEAM search strategy to identify the text information to be recognized.We evaluate the text recognition system on the Street View Text data set, and have achieved a higher correct rate.