Methods and apparatus for text detection

Abstract

A text detection technique comprises local ramp detection, identification of intensity troughs (candidate text strokes), determination of stroke width, preliminary detection of text based on contrast and stroke width, and a consistency check.

1 . A method having scanned intensity information as input for detecting text in a scanned page by observing a very strong contrast in a localized region between a dark side and a light side, the method comprising: 
 determining a stroke width;    contrast-based text detection processing;    wherein the localized region comprises a substantially sharp edge between the dark side and the light side; and    whereby any of black text on white background, black text on color background, and white or light text on a dark background are detected.    
   
   
       2 . The method of  claim 1 , further comprising measuring a color saturation value and using the value to improve detection accuracy, wherein the color saturation value of the dark side is required to be small.  
   
   
       3 . The method of  claim 2 , further comprising preliminarily single pixel processing to estimate the color saturation value using prior color information provided by the scanner.  
   
   
       4 . The method of  claim 1 , furthering comprising detecting the presence of half-tone pixels by using a local indicator to improve detection accuracy.  
   
   
       5 . The method of  claim 4 , wherein the half-tone detection is obtained through an algorithm for half-tone detection.  
   
   
       6 . The method of  claim 1 , wherein the pre-processing further comprises: 
 detecting a local ramp; and    identifying an intensity trough.    
   
   
       7 . The method of  claim 1 , wherein the contrast-based text detection processing further comprises: 
 detecting text preliminarily based on local contrast and stroke width; and    consistency checking.    
   
   
       8 . The method of  claim 1 , wherein the observing a strong contrast further comprises: 
 detecting text preliminarily based on local contrast; and    consistency checking.    
   
   
       9 . The method of  claim 6 , further comprising: 
 detecting a local ramp;    identifying an intensity trough;    detecting text preliminarily based on contrast and stroke width; and    consistency checking.    
   
   
       10 . The method of  claim 6 , wherein identifying an intensity trough uses a finite state machine algorithm, the algorithm having a sweeping procedure.  
   
   
       11 . The method of  claim 6 , wherein the stroke width determination step further comprises: 
 determining a width and a skeleton, wherein the width is a distance value and the skeleton is a skeletal line; and    detecting closely touching text strokes.    
   
   
       12 . The method of  claim 11 , wherein the width and skeleton determining step further comprises: 
 setting the width value to the smaller of a vertical distance and a horizontal distance between two edges of the stroke; and    determining the skeletal line as a roughly equidistant line from the edges.    
   
   
       13 . The method of  claim 11 , wherein the detecting closely touching text strokes further comprises detecting a pattern of dark-light-dark (DLD) in a horizontal or a vertical direction within a very small window.  
   
   
       14 . The method of  claim 7 , wherein the detecting text further comprises deciding whether a current pixel is a text pixel by using the local contrast present in an N×N window having a center over a set of pixels and centered at the current pixel, and stroke width at the current pixel.  
   
   
       15 . The method of  claim 14 , wherein N=9.  
   
   
       16 . The method of  claim 14 , wherein numerous statistics of the pixels within the N×N window are collected by using a set of thresholds.  
   
   
       17 . The method of  claim 16 , wherein 
 the set of thresholds comprises any of: 
 a first minimum intensity level for text background;  
 a maximum intensity level of text to be detected;  
 a second minimum intensity level for text background around crowded text strokes, wherein the second minimum intensity level is smaller than the first minimum intensity level;  
 a medium threshold value, wherein the medium threshold value is around 50% intensity;  
 a first maximum width of a stroke, wherein the first width is considered thin; and  
 a second maximum width of a stroke, wherein the second width is considered very thin; and  
   wherein the numerous statistics comprise any of: 
 a number of pixels that are thin;  
 a number of pixels in the center of a 3×3 window that are thin;  
 a number of pixels on a skeleton, wherein the skeleton pixels are very thin;  
 a minimum width among pixels of the center 3×3 pixels;  
 a second smallest width among the pixels of the center 3×3 pixels, wherein the second smallest width is equal to the minimum width among pixels of the center 3×3 pixels if more than 1 pixel has the minimum width;  
 a highest intensity present in the N×N window;  
 a number of light pixels;  
 a number of non-light pixels;  
 a number of non-light pixels detected as half-toned from a half-tone detection module;  
 a number of dark and neutral pixels;  
 a number of dark and colored pixels;  
 a number of colored pixels with medium intensity;  
 a number of dark and neutral pixels after boosting;  
 a number of pixels in the center 3×3 window, wherein the pixels are dark to medium in intensity;  
 a thin flag set to 1 if the stroke is thin, or set to zero otherwise; and  
 a background flag set to 1 if the center 3×3 pixels are all light, or set to zero otherwise.  
   
   
   
       18 . The method of  claim 16 , further comprising determining if the current pixel is in a category of a set of predetermined categories using an associated algorithm and the set of thresholds, wherein the thresholds are chosen empirically.  
   
   
       19 . The method of  claim 18 , wherein the predetermined set of categories comprises: 
 Text Outline;    Text Body;    Background; and    Non-text.    
   
   
       20 . The method of  claim 18 , further comprising moving the center of the N×N window by J pixels to obtain a subsampled text tag.  
   
   
       21 . The method of  claim 20 , wherein J=3.  
   
   
       22 . The method of  claim 7 , wherein the consistency checking further comprises: 
 accumulating a set of statistics using an N×N window of text tags and a set of thresholds; and    deciding by using the set of statistics if each of the text tags is any of: 
 Text Outline;  
 Text Body;  
 Background; and  
 Non-text.  
   
   
   
       23 . The method of  claim 22 , wherein the N×N window further comprises N×N blocks, each block representing J×J pixels.  
   
   
       24 . The method of  claim 23 , wherein N=5 and J=3.  
   
   
       25 . The method of  claim 22 , wherein set the of thresholds comprises a maximum number of Non-text blocks threshold.  
   
   
       26 . An apparatus for receiving scanned intensity information as input for detecting text in a scanned page by observing a very strong contrast in a localized region between a dark side and a light side, the apparatus comprising: 
 a module for pre-processing for stroke width determination; and    a module for contrast-based text detection processing;    wherein the localized region comprises a substantially sharp edge between the dark side and the light side; and    whereby any of black text on white background, black text on color background, and white or light text on a dark background are detected.