Research On Text Location In Natural Scenes Using Edge Detection And Texture Analysis

One important part of digital image processing is getting text information accurately in natural scenes. The images include road signs, shop signs, bill-boards and posters, etc. That contains lots of useful information for people. Reading the text from natural scenes is a very difficult task. The primary challenge lies in the variety of text: it can vary in font, size, orientation, and position. Text can also be blurred from motion or occluded by other objects. The first step is text locating, then is character recognition and word recognition.An approach based on edge detection and texture analysis is presented. We normalize pixels value between 0 and 255 to effectively handle luminance variations of the picture. The text is designed with high contrast to its background in both color and intensity images. So we can use an edge detector to obtain the edge set, and then get candidate text regions. A slip window was used to obtain local features of candidate text regions. Before extracting features, homogeneity mapping were used to enhance text. Grey level co-occurrence feature is one kind of popular texture features for texture analysis. Statistical features are also used in our research.SVM (Support Vector Machine) is an import technique for data classification. We briefly introduce SVM basics which are necessary for explaining our procedure. We choose RBF kernel among four basic kernels, then the penalty parameter and kernel parameters are chosen. We recommend cross-validation and grid-search to find the best parameters, then use the best parameter to train the whole training set. The cross-validation procedure can prevent the overfitting problem. SVM is an effective classification method, but it does not directly obtain the feature importance. At first, statistical analysis of the features to determine which image features are reliable. Then using various feature selection strategies to select and reduce the dimensionality of the feature space. A classification task usually involves with training and testing data which consist of some data instances. Image smoothing postprocess is needed on testing images to get the final text area. For text locating it is unrealistic to expect a system to agree exactly with the bounding rectangle for a word identified by a human tagger. So we use two different evaluation system, one was proposed by ICDAR, the other was designed by us.The results show that the proposed method can deal with different size text and have a good accuracy.