Application Research On Text Location In Image Based On Maximally Stable Extremal Regions

With the emergence of many mobile devices, the number of natural scene images captured by mobile devices become explosive growth. Text localization in natural scenes is a necessary step for the content analysis in natural scene images. There are great prospects in many systems, such as automatic vehicle driving system, library retrieval classification system, instant text recognition, guiding the blind navigation systems. In contrast to the document text recognition, text localization has less effectivity in natural scenes because of the complexity of environments, flexible access to images and text style diversity. Therefore, text localization and recognition become the research hotspot in the fields of image processing and pattern recognition. The thesis focuses on the methods of the text localization in natural scenes.This thesis detailedly describes the features adopted by text localization, including color features, point and area edge features, texture features, stroke width transform and mixing characteristics. At the same time, an overview on pattern recognition algorithms adopted by text localization is also given. The maximally stable extremal regions(MSER), including its affine invariance, stability and low computation complexity, is elaborated.In the application aspect, the text localization is based on the extraction of MSERs. A best operator(Canny operator) for edge enhancement MSER areas by comparing the existing edge operators(Canny, Sobel, Laplacian, Robert and Prewitt) is chosen. This method is used for character segmentation due to blurred pixels connected, cutting connected pixels in the characters’ holes. Then, we apply the stroke width transform improved by distance transform and geometric constraints. Finally, the text areas in the text candidate regions can be located. This application is used in actual projects successfully.In another application, the MSER areas combined with convolution neural network for classifying the text regions, is adopted. The pretreated the MSER areas are as the input of convolution neural networks. A best structure of convolution neural networks in characters classification is used in this application. According to the results of experimental classification, this method achieves better results, which overcomes the computational complexity caused by the extensive use of sliding windows. Therefore, this method demonstrates the application potential of MSER areas combined with the convolution neutral networks in text localization.