Unsupervised Text Segmentation Using Color and Wavelet Features

Abstract

AbstractSince the number of digital multimedia libraries is growing rapidly, the need to efficiently index, browse and retrieve this information is also increased. In this context, text appearing in images represents an important entity for indexing and retrieval purposes. Often, text is superimposed over complex image background and its recognition by a commercial optical character recognition (OCR) engine is difficult. Thus, there is the need for a text segmentation process, including background removal and binarization, in order to achieve a satisfactory recognition rate by OCR. In this paper, an unsupervised learning method for text segmentation in images with complex backgrounds is presented. First, the color of the text and background is determined based on a color quantizer. Then, the pixel color and the standard deviation of the wavelet transformed image are used to distinguish between text and non-text pixels. To classify pixels into text and background, a slightly modified k-means algorithm is applied which is used to produce a binarized text image. The segmentation result is fed into a commercial OCR software to investigate the segmentation quality. The performance of our approach is demonstrated by presenting experimental results for a set of video frames.