Research Of Scene Text Recognition

With the development of smart devices,researchers is paying more and more attention to scene images which are captured by smart devices and high-level semantic information contained in these images.Optical character recog-nition(OCR)perform poorly on scene text detection and recognition,for the complexity nature of scene images.One reason is that scene images are always taken under uncontrollable condition,which cause problems such as low reso-lution,illumination,blur and etc.The other reason dues to text in scene images varies in font,size and color.So new solution is needed for recognizing scene text.That's the backgournd of this thesis.The main work of this thesis included:(1)Research and analyze state-of-the-art of scene text recognition.(2)Implement an end-to-end scene text recog-nition system-MatE2E system-based on convolutional neural netwcrk(CNN).The system learns character features based on CNN,and trained two classifiers to classify characters.There are two main modules in the system.One is text detecting module.It combins trained CNN classifier and use sliding window to detect text.The other is text recognizing module.It works the same way as text detecting module but used another trained classifier to recognize character.To improve recognization precision,the system used a lexicon to correct rec-ognization result.(3)The MatE2E system is validated on ICDAR 2011,2015 and street view text(SVT)datasets.The results shows that the system is able to recognize scene text with high precision.The validation proves that the MatE2E system has reference value to real world application.Though it has limitations like it is unable to recognize text other than English letters and digits,and it also needs to be improved on speed of recognization and rotated text detection and so on.