Document Text Detection Algorithm

Compared with the traditional document image obtained by the scanner,the document image captured by the mobile phone is often uncertain because of the shooting scene and the angle,document image in the background,the direction of the text,lighting and other aspects are often uncertain,and the document itself exists back handwriting Infiltration,weak contrast text and other degradation of the situation,the existence of many phenomena to the document text localization algorithm has brought great difficulties.Around this issue,this paper improves the FASText text detector,the main results are as follows:(1)A binarization algorithm of document image based on contrast enhancement and background estimation is proposed to solve the problem of text degradation and the selection of the FASText key point threshold m.The algorithm uses the different contrast enhancement algorithm to enhance the document image according to the difference of the luminance information of the different sub-regions.The background of the document image is removed by processing the enhanced document image by using a morphological closed operation,and the binarized image of the document image is obtained by combining the Ostu and Sauvola algorithms.The broken handwriting repair is performed by the distribution of the neighborhood of the edge pixel's gray value.The experimental results show that the F-Measure(FM)is 90.47%,the peak signal-to-noise ratio(PSNR)is 19.15dB,the negative rate index(NRM)is as low as 0.0538,and the classification error measure(MPM)is reduced to 0.00033.The validity of the binarization algorithm proposed in this paper is verified.(2)This paper solve the problem of repeated text detection,text localization and non-text area introduction of FASText text detector.Firstly,the distribution of FASText key points and color attributes are used to filter out pseudo key points.Secondly,the document binary image is morphologically processed and the connected region is extracted by the scanning line seed filling algorithm.The candidate text area of the document image is obtained by the analysis of the connected area.Finally,the non-maximum suppression is used to eliminate the text repetition detection phenomenon.The experimental results show that compared with the FASText text detector,the text repetition detection rate is reduced from 4.2 to 1.7,and the text leakage rate is reduced from 10.3%to 7.9%.(3)A double threshold classifier is designed.The classifier is used to classify the candidate text set into strong text set,weak text set and non-text set.Using the characteristics of similar text attributes between adjacent text,In the weak text,we find the text with similar text attributes with the strong text set,and add the strong text set to the final strong text set.The experimental results show that the recall rate of the proposed document text detection algorithm is 85.1%,the accuracy rate is 87.6%,and the F-Measure is as high as 86.3%,which verifies the effectiveness of the proposed algorithm.