Text Detection In Natural Scenes

Natural scene is not only rich of image information,but also contains a lot of text information.Scene text is a key clue to understanding the content of the scene.The localization of scene text from natural scene images is an important and challenging research problem in the computer vision.The complex background,different illumination condition,variation of the text in terms of characters color,lightness,font,size,interval,contrast,alignment and texture make the problem of text detection very difficult.And the factor such as perspective distortion caused by camera also increases the difficulty of text detection.But,the significance of text detection is very great,such as image retrieval,video retrieval,image notation,Web search and image understanding.Text detection is a very important step for text identification,the results of text detection directly affect the results of the whole system.This thesis presents a study on text detection in natural scene based on the research of scholars both at home and abroad.The proposed method combines maximally stable extremal region(MSER)algorithm and stroke width transform(SWT)algorithm,and it is efficient for text detection in natural scene.We extract connected components in images by using MSER algorithm.An AdaBoost classifier is trained to determines the adjacency relationship and cluster connected components to get candidate text regions by using their pairwise relations.As text region consist of a set of pixels which have similar stroke width,SWT operator is performed for every candidate text region.According to the feature of stroke width,a non-text filter is designed to verify candidate text boxes.Finally,we get the rectangle areas of text regions in the image.ICDAR 2003 dataset and protocol are used for text detection evaluation in this thesis and we compare our algorithm with other text detection algorithm.According to the experiment results of this thesis,the presented algorithm can locate the text region in the image accurately,and this algorithm can be very valuable in application.