Research And Implementation On Text Recognition In Image With Complex Scenes Using Local Features

Problem of text recognition and reading in image is vital and funda-mental in digital image processing and computer vision. Compared with scanned images, camera-based nature and complex scenes have more va-riability including: a) text fonts and character thickness; b) text orienta-tion and alignment; c) background as well as foreground color and texture; d) camera perspective and geometric deformation; e) low image solution, non-uniform illumination and noise. All these factors combine to make the problem a typical kind of object recognition rather than optical cha-racter recognition (OCR). OCR has very strict requirements and restric-tions of input on structural and normative aspects, which makes OCR-based framework a limitation. Efforts can be put upon the location and preprocessing steps to provide well structured and standardized input for the following OCR component, while it is a hardship to process and improve these areas better.Different from OCR-based framework, present proposes a Lo-cal-Feature-Based text recognition framework. This framework applies the technologies in content based image retrieval; builds template charac-ter images datasets; and recognizes the text of input scene through local feature matching and pattern recognition among the datasets. In this paper, two kinds of recognition systems are implemented. One is based Bag-of-Words Model and the other is based Point-to-Point Matching. Both of them are typical application of local feature in image and video processing. Compared with OCR-based framework, local feature based has such superiorities, i) getting rid of some preprocessing steps, for ex-ample enhancement, binarization, layer analysis and geometric normali-zation; ii) by utilizing geometric and photometric invariant local feature and applying targeted voting and geometric consistency constraint, over-coming the limitation of text orientation, irregular alignment, non-uniform image solution, camera perspective and deformation within OCR-based framework; iii) through building multi-languages and mul-ti-fonts template character images datasets, realizing roust recognition in languages, fonts, types and styles.Experiments are mainly focused on 5 kinds of languages including Chinese, Japanese, Korean, English and Arabic. Test results show that proposed method and architecture gives satisfactory recognition accuracy and performance under complex scenes.