SWT voting-based color reduction method for detecting text in natural scene images
ELCVIA Electronic Letters on Computer Vision and Image Analysis
In our PhD thesis  we give a very detailed and in-depth survey of natural scene text detection methods and propose two novel methods, namely SWT (Stroke Width Transform) voting-based color reduction method  and SWT direction determination method  . SWT voting-based color reduction method (to which we will refer also as SWT-V) is a novel text detection method that -opposed to many other text detection methods -combines both structural and color information in order to detect text. The
... detect text. The proposed method upgrades the text detection oriented color reduction method (to which we will refer to as TOCR)  with the additional SWT voting stage and substantially outperforms other state-of-the-art text detection methods. All the image colors rich with SWT pixels  that most likely belong to text characters are blocked from being mean-shifted away in the color reduction process. One of the disadvantages of the SWT method , however, is the problem of 'light text on the dark background' described in the following paragraphs. To cope with the problem and in order to provide true SWT values to the SWT voting stage we propose an adaptive SWT direction determination method. The method uses SWT profiles to partition an image into subblocks and analyses their SWT histograms of both SWT search directions  . Text detection literature does not explicitly address the SWT direction issue, therefore, the proposed method represents a unique scientific contribution to the research field. All text detection methods were evaluated on the CVL OCR DB text detection evaluation dataset [2, 5] . The main disadvantage of the TOCR method  stems from the fact that non-dominant text colors often vanish in the color reduction stage. When a certain text color lies too close to a more dominant background color in the RGB color space it is mean-shifted away from its original position. A typical example of vanishing text color is depicted in Figure 1b . A blue 'Clairefontaine' color in Figure 1a is lost in the color reduction stage. To avoid such a problematic behavior we propose the SWT voting-based color reduction method. The proposed method upgrades the TOCR method with the additional stages (depicted as yellow rectangles in Figure 2 ), namely 'SWT filter', 'SWT direction determination' and 'SWT voting'. In the 'SWT filter' stage we create a SWT image  and obtain its correct SWT directions in the 'SWT direction determination' stage. Since the non-zero pixels in the SWT image most likely belong to the text regions in the image  they serve as a very useful clue in the subsequent 'SWT voting' stage.