Research Of The Text Information Extraction In Images With Complicated Background

Text data in image play a significant role in the intelligent control system and information detection and retrieval system because they contain plenty of valuable information. More and more information appear as image or video with the increasing application of mobile and portable image capturing devices. It is urgent to endow computer with the abilities of processing, identifying and understanding text information in images. However, computer’s capability of recognizing and understanding the text information cannot meet the requirements of the practical application. Researchers have paid their attentions on efficiently extracting the text information for long time. Especially, the extraction of the text information from the complicated background is still an open issue.This paper focuses on extracting and locating the text information contained in the images with complicated backgrounds. The images are captured by portable devices. Analyzing the intrinsic property of the text-in the scene images, we propose three methods for extracting and locating text in scene images, and implement a text extraction and recognition system. The contributions of this paper are as follows:(1) A text locating method based on texture and statistic features is proposed. The simplified mean shift algorithm is used to smooth the input image, which can remove the noise and reserve the strong detail information of the image. The edge map of the image is obtained to extract the stroke feature to filter non text pixels, while the statistic features of the blocks are employed to filter non text blocks. Experiments show that this method is fast with high recall. It is efficient for detecting the characters which adhere to the complicated background.(2) A text locating method based a modified visual attention model is presented. Itti’s visual attention model is modified:First, the Gaussian pyramid can adjust its number of the layer according to the size of the input image; Second, the intense conspicuity is applied to be the saliency map, and its normalization is left out in case of weak text regions; Third, the saliency map, with the same size as the input image, is yielded by up-sampling the feature maps to stand out the detail information. The connected components are filtered by the Histogram of Oriented Gradient feature. The saliency map is then employed to filer the non-text regions. Experiments demonstrate that the method is capable of detecting the text regions with low contrast and achieving good performance.(3) A text locating method based on the sale space is proposed. The four templates of Sobel edge detector are modified by replacing the two diagonal Sobel templates with Ridge templates to avoid the adhesion of the text with the background. The property of the strongest responses is used to filter the non text regions. The responses, locating on the crosses or the endpoints of the strokes, always correspond to the stroke width of the character. Therefore, the stroke width of the candidate regions is applied to be the heuristic condition to search the related stronger responses in the scale space. The non text regions can be filtered by the distribution of the response. Experimental results prove that this method has high precision.(4) An automatic text detection and recognition system of the natural scene images is implemented. The text locating method is used to obtain the text regions of the input natural scene image. Then, the scale normalization is applied on the detected text region after being binarized. Finally, an OCR package is employed to extract the text information of the detected text regions.(5)The performances of the above three methods are compared and discussed. Experiments indicate that each method has advantages and disadvantages. The method based on the texture and statistic features is efficient for the images with complicated background although its performance is worse than the other two; The method based on modified visual attention model can successfully detect the low contrast text regions in spite of the fact that it is a little worse performance compared with the method based on the scale space. The method based on the scale space is capable of removing the text like background, with the highest whole performance, especially the precision.