Korean Text Extraction by Local Color Quantization and K-means Clustering in Natural Scene

Understanding of text information in a natural scene is very useful and important in many real applications. Extracting text information from signboard in natural scenes is involved to variations of existing problems, such as uneven illumination and shadow, blurring and scratching, multi color, and complex background. Many approaches have been proposed to over challenges in text extraction. However, simple approaches based on grayscale image cannot dealt with difficult conditions in natural scenes; complicated ones based on color information give burden to real system. In this paper, an effective method for the Korean text extraction from signboard images is proposed, which can handle degradation condition in natural scenes. The proposed approach is based on local color quantization and k-means clustering on separate characters. Color quantization gives solution to problem of color text segmentation method with low computational complexity and good performance of quality. First the detected text region is separated into local characters with relatively uniform illumination and background and using 3-means clustering with cosine similarity has been applied to segment text from the background. Natural images from the test database, collected from mobile devices, are used in the experiment and the results show the performance of the proposed method.