Text extraction in real scene images on planar planes

This paper proposes a hybrid approach of a texture-based method and a connected component-based one for extracting texts in real scene images. For detecting texts having a lot of variations in size, shape, etc, we use a multiple-continuously adaptive mean shift algorithm on the text probability image produced by a multi-layer perceptron. It is assumed that the scene text lies on planar rectangular surfaces with homogeneous background colors. We correct perspective distortion using warping parameters calculated after segmentation of an input image. We can detect and reconstruct text images accurately and efficiently.