Improved Image Segmentation Methods For Superpixel And Graph Cuts

Large numbers of text information in natural images, such as traffic signs,billboards, business signs, street signs, which provide direct and key cues for scene image understanding and analysis. Consequently, it is important to develop a kind of automatic identification text information tools for text retrieval and analysis. It can be widely used in the fields that need to image retrieval, intelligent monitor, guiding for the blind navigation and other fields.To realize the automatic recognition of text cannot do without the image of text segmentation. Image segmentation is a key step in the field of computer vision. The image is easily affected by the shooting angle and light which can be deformed,blurred, broken and so on. Traditional document segmentation techniques are not effective in the application of complex natural scene text segmentation because of the background is more complex. Recently, more efficient text segmentation in complex natural scenes is still one of the hottest research focuses in the field of computer vision and document analysis.In this paper, the segmentation of text images in complex background is studied,the main work is as follows:Inspired by the concept of super pixels, in order to reduce the redundancy of the image local information and improve the efficiency of the graph model, this paper constructs graph model based on the superpixel regions instead of pixels. So, this paper introduces the development background and history of super pixel. Then, this paper chooses the suitable method of SLIC through combing various super pixel generation methods and evaluating the advantages and disadvantages of the application of various scenes. Finally, we improve the super pixels generation algorithm based on SLIC appropriately.Image segmentation step based on graph cuts is discussed in detail. In this paper we put forward a kind of scene text segmentation algorithm based on improved super pixels and graph cuts. The algorithm uses super pixels instead of pixels to construct agraph model. The text and the background model are constructed by using the GMMs model and the SVM based posterior probability model, and the matching degree between each local region and the model is introduced into the graph cut model to realize the calculation of the likelihood.Finally, we carry out the experimental test in the Matlab software environment using the laboratory database. we use optical character recognition software(OCR) to test the recognition effect of segmented image.The results show that: this algorithm has good performance and certain application value.