Research On Text Location Technologyin The Image

Text localization in image is a prerequisite for many image processing applications, such as image retrieval system, intelligent transpo rtation system, license plate recognition system and so on. Although a certain amount of development was achieved on this technology, the interference of complex background in images are affecting the effects of text localization algorithm seriously, and the localization accuracy remains to be improved. How to improve the anti- interference performance of text localization algorithm with basing on feature fusion and improving is the main problem for this paper to solve.This topic research task is rooted in the National 863 Planned project “Telecommunications Network Security System”, and the text localization approaches of overlay-texts and scene-texts were studied. According to the project requirements, we designed a text extraction system in images.In this paper, the research content and contributions are as follows:(1) An overlay text localization algorithm based on bidirectional maximum gradient difference was proposed in this paper. The distinction between text and the background was improved and the localization accuracy of overlay text in the complex background was effectively increased by our algorithm through extracting the horizontal and vertical maximum gradient difference features. Firstly, the localization of the candidate text regions based on proposed bidirectional maximum gradient difference feature was realized. Secondly, text regions were excluded based on a priori knowledge. Finally, the crossing line features were integrated with the aspect ratio to achieve accurate positioning of the text area. Experimental results showed that the bidirectional maximum gradient difference feature can made full use of gradient information, which better enhance the distinction between text and background and the positioning accuracy of overlay text in complex background images. Thus when the problem of overlay text in complex background and the processing speed were resolved at the expense of some of the cases, a better localization result comparing with the existing algorithms was achieved by our algorithm.(2) A scene text localization algorithm based on the fusion of MSER(Maximally Stable Extremal Region) and visual saliency is proposed in this paper. From the perspective of multiple features integration, the complimentary of MSER and visual saliency is employed to effectively improve the single MSER robustness to the background. Firstly, MSER features and visual saliency features are extracted. Secondly, fusion judging strategies were used to exclude MSER with low saliency to achieve the screening of candidate characters MSER. Thirdly, ellipse fitting of MSER was brought in and new character feature was reconstructed to exclude pseudo-character MSER. Finally, the left character MSERs were dealt with bidirectional projection combining to realize text localization. Simulation results showed that the introduction of visual saliency effectively improved the immunity of MSER characteristics against complex background, thus the accuracy of this algorithm on scene text localization was improved in complex background images.(3) Two localization algorithms mentioned above are combined with the Tesseract-OCR technology to achieve a text extraction in images system. The application effect on the current network data showed that the text localization algorithm proposed in this paper achieved relatively better availability and the system’s recall rate reached 56.2% and precision rate 64.8%.