Research On Scene Text Recognition Method And Software Realization

Information exchanges play a very important role in our daily life, and text information can be found everywhere in the real environment. It is very important to obtain text information from scene images automatically, and many text recognition methods have been proposed these years. Optical Character Recognition(OCR) is one of the many text recognition methods with the highest accuracy rate and lowest run time, however, there is one shortcoming of OCR: it can only process images with single background. In most applications, images that need to be recognized are usually captured from a natural scene, which means these image backgrounds are complicated and unpredictable. Naturally, the technology of text recognition in natural images has developed rapidly.English text recognition in natural images has been applied widely due to the popularity of the language itself. Image preprocessing, feature extraction and results refining are issues in English text recognition worthy of further study most. Based on consideration of those issues and other genius works done recently, we focus on preprocessing and results refining. The main works are as follows:1. We reclassify all characters. The classic method labels all the characters into 62 categories according to senior semantic information. By calculating the probability of misclassification between every two categories, we merge the category pair with a high misclassification probability. Our new classification approach raises classification accuracy via reducing total amount of labels.2. We preprocess an input image using GrabCut. We propose the calculation of separate confidence(SC) of every column by taking the results of GrabCut into consideration. After computing all SCs, we assign a start score to each sliding window.3. We define a method to refine the recognition results based on the character number in the word, we use large lexicon information in the refining.In our method, we raise the accuracy of character recognition based on our new classification of characters, and we also refine text recognition using large lexicon information.