Research On Text Location And Recognition In Natural Scenes Of Image

Detecting and recognizing text in natural scenes, such as sign boards on streets and buildings, traffic signs, movie marques and so on, is a core part of computer vision applications, roboticsand text reading programs for visually impaired person. Recently, with the rapid development of computer vision and robot vision and artificial intelligence, how to extract text information from scene images with high efficiencyhas become one of the hot research topics. Detecting and recognizing text includes three steps:character detection, text line detection and text recognition. On the basis of studying and summarizing existing excellent algorithms and according to the characteristics of text in natural scenes, in this thesis, text detection, text line detection and text recognition under complex scenes are deeply studied. The main contents are as follows:(1)Two kinds of character region detection methods are designed for different application scenarios.In the condition of image blur, unven illumination and low light, etc., the edge portion of text will have a serious distortion, so that it is difficult for existing text detection algorithms to achieve desired results. As to solve this problem,a character detection method byusing binarization is proposed, which usesmultistage binarizationandtwo-pass connected-component labeling algorithms to detect candidate character region sin three channelsof image and gray image.Considering that compared to background texts always have a clear outline in most natural scenes.In order to improve the detection speed of this condition, a candidate character regions locating method based on image layeringis proposed. The image is viewed as a tree model which is stacked by a series of components, and character areasare detected by component detection. And part of non character areasare removed based on the structure characteristic of the image tree.(2) In order to find candidate text lines in character regions set and to solve the problem of detecting text lines with random directions, a text line detection method based on graph model is proposed. The graph model is constructed by the relation of regional spatial distribution information and geometric information, and the search path is prunedbased on graph search strategy. In order to remove non text lines, the joint characteristics of regions in the candidate text lines are extracted, andoverall features with adaptive adjustment according to the inclination direction are extracted. The result proves that the method performs well for text lines with random directions.(3) The texts are recognized by the method combining Tesscract OCR and Google spell checking. Due to the complexity of the text in the natural scene, it is difficult for Tesseract OCR to obtainideal result. In order to solve this problem, firstly, textimages are corrected before using Tesseract OCR, and obstacles in the background are removed by combining with the color distribution offoregroundand boundingbox of text region. After that tilt correction and shear correction are operated on these images. Then the recognization results are modified by Google spell checking. The result shows that the method improves the recognition accuracy.