Research On Chinese Character Recog- Nition In Natural Scene

With the generalization of intelligent terminals and rapid development of internet technology, images and videos are getting more popular among people as the media for transferring information. Consequently, people's demands for disposing numerous images also become increasingly urgent. The issue how to quickly and precisely extract the useful information out of images by computers becomes a hotspot for many scholars'research. The character recognition technology for natural scenes has extensive application prospects, such as, the content-based image retrieval (CBIR), intelligent transportation system, visual aids system and pilotless automobile and so on. At present, the studies at home and abroad are mainly concentrated on the recognition of English letters in natural scenes, while the researches on recognition of Chinese characters are still lacking. In this paper, on the basis of a comprehensive research on the recognition of English letters, an appropriate method for recognizing Chinese characters in natural scenes will be designed and implemented.Firstly, the algorithm of maximally stable extremal regions will be combined with mathematical morphology to serve as the pick-up algorithm of candidate character zones and adapt it to the detached strokes of Chinese characters. Through experimental comparisons, the influence of the pick-up algorithm parameters is verified.Secondly, a set of heuristic filtering rule based on geometrical characteristics is designed according to the features of characters, during which the candidate character zone is handled with ellipse fitting, as the eccentricity ratio of the fitting ellipse is regarded as one of the standards to confirm the character zone. With the feature of rapid calculation, a preliminary identification for text area and non-text area is achieved. The experiment shows that this heuristic filtering rule can filter out most non-character zones on the basis of maintaining the original character zone, which greatly enhances the efficiency of the whole algorithm.In this paper, the histogram of oriented gradients features of a powerful feature descriptor with illumination robustness is selected to be the eigenvector of character zones. By training a support vector machine classifier, the text areas are precisely detected and located. It is well-known that Chinese characters are numerous with complicated structures and variable typefaces; thus, it's very difficult to collect Chinese characters covering all classes from natural scenes. As a result, this paper will study various common Chinese characters in the natural scenes and create Chinese characters of different scales for every kind of typeface to be samples instead of directly adopting the Chinese characters in natural scenes; then make recognition through KNN classifier. As the experiment shows, the method adopted in this paper results in a satisfactory results.