Design And Realization Of Text Extraction From Scene Image Based On Embeded Platform

Text embedded in the natural scene images is an integral part of the images’semantic content. Scene text exists in many forms such as traffic signs, subtitles, store brand, text banners, building plaque. This kind of information can provide valuable clues for scene image understanding. Extraction and recognition of text regions from natural scene images will benefit for data retrieval, query, understanding, compression of the images, which helps improving the management efficiency for the image data.Traditional document analysis technology such as OCR is mature, and has great achievements in the last decades, for it can process artificial text accurately and efficiently. However, it is still unable to effectively process natural text character in a scene image. Now, research on natural image process stays in the stage of algorithm design.In this thesis, we design and realize a software package, which consists of two modules:image acquisition module and text extraction module. In image acquisition module, we use HD webcam to capture natural scene image, control the camera via an open-source video processing library based Directshow technology, and then implement an operation-friendly graphical interface via Qt library. Text extraction module is the core of the whole package, in this module, we merge the results of the gray statistical method and multi-scale edge detection method, thus we can precisely locate the text regions from natural scene image and achieve a high correct detection rate. Additionally, this module is perfectly transported into embedded system platforms and performs well. Finally, the thesis adopt an image enhancement method to improve the correct detection rate and does some exploration on character image threshold via filling method and extract the text on line.The realization of this package is based on the C++programming language and Opencv image process library. The software can locate text regions from scene image accurately, and the method used is robust with respect to font, sizes, orientations, and alignment. Experiment shows that the time cost for processing a image with640X480resolution is less than5seconds on Windows platform, which indicates that the software can be applied for real-time applications. Meanwhile, on embedded platform, the software performs well but cost much time, up to about one minute.