Research On Preprocessing Of Optical Character Recognition For Mobile Devices

With the dramatic development of portable digital imaging devices with high performance and low priced and the deployment of high efficiency and high computing power of microprocessors and graphic chips on the platform of mobile devices, as well as the large-scale coverage of mobile Internet, it makes mobile smart devices become an important part of our daily life, which also stimulate the development of applications and programs based on the computer vision and image recognition algorithms which have sprung up in recent years. However, the portable digital imaging devices are generally in complex environments. The results of image preprocessing stage can largely influence the final recognition accuracy rate. This study thus focus on the design and research on the techniques and algorithms involved in the document image preprocessing stage.Firstly, in order to get a clear clue of the development of related researches, this article will review and summarize the key issues, as well as their research achievements when deploying document image recognition algorithms onto mobile platform. Specifically, this research will put more attention on the light balancing and slant and skew corrections for document images. At the meanwhile, a brief summary of feature extraction and some state-of-the-arts machine learning algorithms will be provided. Secondly, we proposed a new light balancing method for the enhancement of under-exposure document image taken by smart phone, in which mathematical morphology operation and a global light balancing equation will be involved. By the comparison of OCR accuracy rate before and after using the new method, as well as other methods, it proofs that the new method can reconstruct and restore the optical character region efficiently which are degraded by shadows and noises under low illumination condition collected by mobile smart phone. In addition, with respect to the slant and skew correction, a strategy by combining connected component analysis and Radon transform will be provided for solving the slant correction. And two global scheme for skew correction will be discussed, in which techniques including Hough line detection and least-square linear fit, as well as connected component analysis and perspective coordinate transform will be involved. Finally, the experimental results illustrate the feasibility and effectiveness of these new algorithms.