Character-Localization in DCT-Compressed Domain
HUANG Xiang lin, SHEN Lan sun (Signal and Information Processing Lab, Beijing Polytechnic University, Beijing 100022)
Segmenting character regions in an image is very important because these characters contain clear clues of retrieving and browsing images from video/image databases efficiently and effectively. In this paper, We propose a method to locate character regions of video/image in DCT compressed domain directly. With the distinguishing characteristics of character's texture (such as horizontal lines, vertical lines, or slant lines in a character) that can be extracted directly in DCT compressed domain, the character regions are segmented from their backgrounds quickly, and the image noises rising during the processing period can be removed by morphological filter. With this method, the compressed bit streams, which are encoded by DCT based encoding algorithm such as JPEG, MPEG 1/2, etc., can be processed directly to locate the character regions in image, just a very small amount of decoding is required (Huffman decoding only). So, the amount of data which want to process is smaller, the processing speed is faster and the demand of computer memory is less. The experimental results show that the correct localization rate of this algorithm is higher.