Text Detection And Recognition In Image

In recent years,detection and recognition of image text continues to expand the application scope in image searching,license plate recognition,fast image documentation,industrial production line and so on,which has attracted great interest among researchers.However,since the complex image background,variations in lighting and angle,as well as the huge number of text languages,the final precision of recognition is difficult to satisfy actual applications.From the view of practical applications,the main work of this paper is to study the text detection and recognition methods for images in specific scenes.With the existing problems of text detection for images,three methods of text detection in different scenes are discussed in this paper.For those illumination affected and background slowly changed images,from the view point of image frequency domain,this paper takes the homomorphic filtering technique to remove the low frequency background signal and reserve the high frequency text information.Meanwhile,the detection of text lines is conducted on the basis of characteristics of abundant edge information in text and the morphology methods as well.For those complex industrial scene images,there are always dependency relationships between texts and objects.Accordingly,text detection is changed to text-related object detection.As a result,false alarm rate can be significantly reduced in this way.This method is proved of high precision and obtained prominent effect in practical industry applications.For those nature scene images,the MSER algorithm based on edge enhancement is presented from the point of improving the quality of region capture.And then,the character sorting tree is build to sort the character region.Next,a multi-level fusion strategy is proposed to detect the multi-direction text.Finally,the random forest classifier is used to evaluate the candidate texts.Experimental result shows that the method can improve the recall rate of text lines and the precision of recognition.The next part is about segmentation and recognition of text lines in images.For those simple or gradual background images,from the view point of practical industrial application,this paper adopt the unsupervised segmentation algorithm to segment single character based on the characteristics of the similarity between Gaussian mixture distribution and vertical projection curve of character region or edge.Then,a CNN(Convolutional Neural Network)model is trained for single character recognition.For those relative complex background text line images,the sliding window method is utilized for recognition.First,a CNN model is used to slide the window from left to right on the text line images for distinguishing characters and non-characters.Next,convex hull is adopted to detect the curve formed by the confidence of sliding window sequence recognition results.Then,a SVM model is used for extracting the width and height features of convex hull to classify characters and non-characters.After the segmentation of characters has been completed,the trained CNN model is conducted for single character recognition.In view of the above mentioned segment-based recognition method are finally both using CNN models to recognize single character,and the context relations of characters are not considered.Therefore,based on the prior character segmentation method,text sequence recognition with RNN(Recurrent Neural Networks)models is further studied in this paper.Experimental result shows that the text recognition rate has improved with additional sequence recognition models.The finally part is about text line sequence recognition.This paper borrows idea basically form speech recognition technology,and proposed a novel method of recognizing text line from the view point of sequence recognition.In this method,the convolutional layer of CNN model is firstly used to extract features for text sequence,and then the features are sent to the LSTM(Long Short-Term Memory)model for training.Two LSTM models are trained for obtaining the forward and backward context relationships of text sequence.While employing offset to get multiple sequences for the purpose of avoiding improper partition when using them.Next,the trained LSTM model is used to evaluate the multiple character sequence recognition results,and the highest score is chosen as the final recognition result.Experimental result shows that the sequence-based recognition method achieves better performance than the segment-based recognition method proposed above.