Text and Picture Segmentation by the Distribution Analysis of Wavelet

Statistical classi cation is an important topic in image processing. Classi cation helps to interpret images, and it can be incorporated into other image processing algorithms, e.g., image compression [1], to improve performance. A particularly interesting type of classi cation is the segmentation of pictures and text. By pictures, we mean continuous-tone images such as photographs. By text, we mean normal text, tables and graphs. One application is the World Wide Web. With the exponentially increasing popularity of the World Wide Web, web information is handled more and more frequently. Since most web pages are a mixture of text and pictures, the segmentation of the two is preferred in many kinds of processing. This paper presents an e cient algorithm to segment text and pictures using wavelet transforms [2]. Wavelet transforms have played important roles in classi cation of texture [3, 4] and abnormalities in medical images [5, 6]. An application of wavelet transforms is the formation of classi cation features by the statistics of wavelet coe cients. The moments of wavelet coe cients are the most commonly used [3, 4, 5, 6]. In our paper, however, we pay direct attention to the distribution pattern of wavelet coe cients and de ne features depending on the shape of the histogram of wavelet coe cients. The classi cation is block-based, i.e., an image is broken into blocks and every block is classi ed separately. The block size is variable and content-based, however, so that the segmentation can be accurate at the boundaries of two types and avoid misclassi cation due to over-localized region analysis.