Text Information Extraction In Colorful Scene Image

As the rapid development of computer science and multimedia technology,the multimedia information,mainly composed of color images,has rapidly become an important general information media.Texts in color images,such as news headlines, aside,and etc.,usually contain much high level semantic information.So it is very helpful for indexing and retrieving images to recognize and analyze those texts automatically extracted from images.On the other hand,it is also significant for image reusing and image intercommunication among different languages to remove texts embedded in images.The conventional approaches are mainly based on statistics or geometry.The advantage and disadvantage of the algorithm of edge detection,texture and region characteristics are analyzed in this paper.The defection of these algorithms makes it hard to detect the semantics which constituted by scenery.In this paper,the text in the images is taken as a kind of scenery not the characters.In this way,some semantics constituted by scenery are taken as a kind of text.After the mathematical and physical models of Mean-Shift are introduced, Mean-Shift algorithm is adopted to complete the intention.Firstly,extracting all color clusters from the original image with the Mean-Shift algorithm.Secondly,translating every color cluster to gray image.After that,The global threshold segment is adopted to segment the text region after the global threshold segment,local threshold segment and dynamic threshold segment are compared.The characters in the text regions share interspaces between the characters. These kinds of spaces can be detected by projection segment as tough in the projection.After detection,the characters are extracted from the text by projection segment.Finally,the characters in the text are recognized by discrete Hopfield neural network(DHNN).The DHNN method performs rapidly,effectively and is small size of its source code.