A Research Of Text Recognition In Natural Scene Images

Currently,the artificial intelligence technology is widely used.Computer vision technology is one of the popular research areas,among which the text recognition technology in the natural scene has also gained a huge leap.Text recognition in the natural scene is different from recognition of a scanned document.Its image background is more diverse,the text direction is not uniform,and the characters of the text can have different sizes,colors,fonts,etc.Obtaining text information in natural scenes has value and prospects in broad applications such as machine translation,license plate recognition and human-computer interaction.This thesis mainly studies the text recognition in natural scenes from deep learning.The main content is organized as follows:1.Researched an algorithm for synthesizing natural scene text images.Due to the limited data or inaccurate labeling of the manually-labelled text database in natural scene,the training of the model will be adversely affected.Therefore,we have studied an algorithm synthesizing the scene text images that are close to the real scene text image.The algorithm can solve the problem that a large amount of pre-training data is needed in the text localization algorithm.Moreover,a larger set of the scene data and more font types are used for synthesizing more diverse text data.2.Researched the Faster R-CNN target detection algorithm.In view of the fact that there are many multiple directions of text in the natural scene,based on Faster R-CNN,the RPN network is improved and extended so that the algorithm can locate text targets in multiple directions easily in the natural scene.The corner detection method is used to find four different types of corners,then the proposal sampling is performed by different types of corners to obtain quadrilateral region proposals.Combined with the subsequent steps of Faster R-CNN,it facilitates and effectively improves the localization of the text targets.3.Researched a sequence-based text recognition algorithm.As the deep network model is prone to cause network degradation and low recognition accuracy in the natural scene,we used the better feature extraction network and replaced the original Bi-LSTM with Res-LSTM during the encoding stage.In the decoding stage,the global attention mechanism depends on all the feature vectors after encoding,which affects the recognition accuracy of long text sequence.The local attention mechanism is applied to replace the global attention mechanism and improve the accuracy of algorithm for text recognition in the natural scene.