Text Detection And Recognition In Images

Text in images contains plenty of information and extracting information in the images plays an important role in understanding the scenes. In recent years, text extraction in scene images effectively promotes the development of content-based image and video retrieval,network security, visual aids and other fields. The current text location algorithm is far from the practical requirements. In this thesis, we analysis the existing algorithms as well as take the characteristics of text in imagery into account, and we put forward an effective framework for text detection and recognition in images.In this paper, we use the following processes to detect the text in images: (1) detect Canny edges of image, (2) perform morphological operations; (3) find rectangle region of the candidates of the text by connected domain analysis , (4) extract the joint features of HOG and LBP in candidate regions , (5) train text classifier, (6) locate text of images with classifier model. The experimental results show that the proposed method can locate the image text region accurately.To distinguish text and non-text, we train text classifiers by two methods, one is convolutional neural network (CNN) and the other is SVM classifier. The results show that the precision of CNN classifier is 5-6% higher than that of SVM.In the process of text recognition, we directly recognize the word images rather than character images. This method is based on the deep learning. Word recognition is achieved by a classification model of the 88172 classes, each of which is an English word. The result shows that this approach considerably reduces the side effect of non-uniform illumination on the image recognition and it can recognize the text in images more accurately.