Abstract
We propose a new global binarization method suited for character strings extraction from a scene image. To obtain a high-quality segmentation result, binarization is a significant step in character extraction process. Since character strings region of sign board has the wide variety of layout and color, it is hard to binarize this category of images by conventional global methods, which are based on statistical information. The proposed method can select a threshold by evaluating the fineness of the binarized image by using the fractal dimension. By evaluating the compactness and stability of the regions of binarized image with the fractal dimension calculated by blanket method, the proposed method achieves fine binarzation results, which are no way inferior to manually and/or using one of promising global methods known as Otsu binarizing method.
