Research On The Algorithm Of Text Location In Natural Scene

As the electronic devices with the camera function becoming more and more popular, a flood of pictures and video files produced every minute. To help people efficiently and accurately obtain text information from pictures, natural character recognition technology develope rapidly. As one of the most important part of character recognition, Nowadays, text location in natural scene images is popular and significant in computer vision.In this paper, on the background of natural scenes, we study the text location method in scene images. We take the character stroke as the based feature, and propose two text location algorithms in natural scenes on the basis of previous studies: The first text location algorithm is based on connected-commonents. Firstly, we transform the original image to the stroke width image, and search through the stroke width image conditionally to obtain candidate character connected-commonents. Then we select better candidate characters’ by heuristic rules.After that, we extract features from connected-commonents, and train SVM(Support Vector Machine) classifier by learning the datasets, using classifier to verify and delete the noncharacter connected-commonents. Finally we merge the character connectedcommonents according to the text feature, and obtain the location information of the text objects.The second text location algorithm is based on the pictorial structure. A new text model is constructed to describe text object using pictorial structure. We abstract the single character and the relationship between neighboring characters. First, we consider the candidate characters as pictorial’s nodes and compute character energy that is based on three characteristics: average angle deviation, non-noise components,and stroke width vector, to represent the probability wether the node is a character; after this, we consider the link between the neighboring characters as pictorial’s edge and define the link energy to represent the probability wether two characters belong to the same text. For every candidate link, we compute link energy based on the similarity of characters. Then combining character energy and link energy, we define text energy to respresent the probability wether the candidate text is the true text, and we get the final position information by appropriate text energy threshold.In addition, in order to make these methods more effective, we give some preprocessing algorithms: we filter the noise and small impurities in scene by using Discontinuity Preserving Smoothing methods; to get more information of text edge we propose a new way about color edge detection; for the text model we construct, when computing three basic features, all edges must be closed, so we design a closed edge detection operator to meet this requirement.At the end of the paper, we compare algorithms proposed with other algorithms in location effect. Then we point out the shortcomings of these algorithms and give the future expectations.