Automatic Detection and Segmentation of Text in Low Quality Thai Sign Images

Abstract

A system for automatic detection and segmentation of text in low quality Thai sign images is presented in this paper. The method is designed as a part of a real-time Thai sign translator system which can be used in many applications. First, an input image is pre-processed to enhance its quality. Secondly, we apply LoG (Laplacian of Gaussian) to the image for edge detection. After edge detection, in the third step, we perform connected component labeling and morphological operations for contour filling. To detect an element in a Thai sentence, we have to set some appropriate ratios and compare them with each closed region to find consonants, vowels, tones and special symbols. Next, we employ 4-line Thai character criteria for layout analysis. Finally, we use GMM (Gaussian mixture model) to represent foreground and background, and perform color segmentation in selected color model. Finally, a method for perspective distortion correction is also performed to prepare the segmented texts be ready for further recognition process. We tested the system on 192 Thai sign images which contain total of 4681 characters. The images were captured from several environments in various lighting conditions. The detection accuracy is 90.22%