Research On Text Detection And Recognition For Screen-rendered Image

Text recognition on screen-rendered text images has an very wide application in blind assisted,automated testing and online dictionaries.However,due to its low resolution,small font size,and low contrast properties,existing text recognition methods face great challenges in Chinese and English text recognition of screen-rendered images.This thesis apply deep learning method to solve the Chinese and English recognition problem from screen-rendered text images.The work of this thesis mainly includes the following two aspects:1)Text detection and recognition based on character segmentation method in screen-rendered images.Aiming at the characteristics of screen rendering text images,this thesis proposed a segmentation-based method using Goog Le Net.This method firstly uses OTSU binarization,dilation,connected domain detection,connected domain fusion,and vertical projection to extract single character from screen-rendered text images,and then the word-width fusion method is applied to correct the error-segmentation character.Finally,a slim Goog Le Net network is designed using four inception-V2 modules.The experimental conducted on the public datasets CIFAR-10 and ICDAR 2013 and screen-rendered text image dataset demonstrate the performance and practicality of the proposed method.2)A segmentation-free end-to-end method for text detection and recognition from screen-rendered images.In the segmentation-based methods,some sticky characters are difficult to be segmented.In order to solve this problem,this thesis proposed a segmentation-free recognition method based on deep residual network,recurrent neural network and connectionist temporal classification.This method firstly uses OTSU binarization,dilation,connected domain detection and connected domain fusion to extract text lines from screen-rendered text images.Then,in order to recognize text with variational length,this method uses deep residual network,recurrent neural network and connected domain detection to construct an text recognition model.The experimental carried out on the public datasets CVL HDS and ORAND-CAR and screen-rendered text image datasets demonstrate the performance and practicality of the proposed method.In this thesis,a deep learning model is used to study the text detection and recognition on screen-rendered text images from both the segmentation-based and segmentation-free recognition,aiming at the low resolution,small font,and low contrast of screen rendering text images.Improvements to traditional methods.The proposed method improves the text recognition method of screen-rendered text images,and has certain reference value for automatic testing,online dictionary and text recognition in natural scenes.