Koichi Ejiri

Fast and accurate skew detection algorithm for a text document or a document with straight lines

Bit-mapped images are becoming more popular in offices. Skew is a major problem for many otherwise promising applications. To remove the skew, we propose a new algorithm that makes use of both printed characters and straight line(s). Lines on a document are decomposed into small segments of black runs. By checking their connectivities, we can easily tell whether those runs are from the same line or not. To remove any bad effect from variation in line width, we sample a number of different x-y coordinates along the black runs, adjacent to white pixels. Those coordinates determine a correlation function which is used to find the correlation value. If the value is close to 1.0, we compute the higher-probability regression coefficient using the same parameters. The algorithm is effective both for horizontal and vertical lines. The coefficients can also be used to align character lines. The rectangles formed by connected black pixel are extracted using two or three different compression ratios. We can tell whether those characters are from the same character line or not, by checking the coordinates of rectangles in multiple compression images.

Weak model-dependent page segmentation and skew correction for processing document images

Presents an algorithm for fast accurate page segmentation that is as far as possible independent of a model for the document page. The method is based on a reduced image data representation that uses bounding rectangles and run length size distributions contained within the bounding rectangles. These rectangles are the basis for the method of skew detection, column identification and merging of text into blocks, and they help achieve accurate page segmentation. The reduced complexity that rectangles offer insures a fast processing time. The method is applied to a broad range of documents found in a typical office environment. Documents are scanned at 400 dpi and stored as binary images.<<ETX>>

Word frequency distribution in Japanese text

Abstract In our last paper (Ejiri 1993), we proposed a new parameter, G = log (N/L)/{log(N)‐1}, where N is the number of words and L the number of different words, which correlates with constraint of the target text represented in ASCII code. We found that the same measure is applicable to Japanese texts which have no clear word segmentation. By statistical analysis of kanji, kata‐kana and alphabetic strings, Japanese texts were found to have a similar distribution as English texts or computer language. We also introduced a “joint entropy” of string[i] and string[j] where the latter string follows the former string after a fixed distance. Here, the distance means the number of words (or defined character strings) between string[i] and string[j]. This entropy is a measure of redundant description (phrases) in the text. A simpler approach using n‐gram frequency was found to be useful to detect errors in a text recognized by OCR (Optical Character Reader).

Composing virtual environment using images of digital camera

We present a new method to make virtual environment by using a sphere for reference. Images taken with digital camera or video camera are projected to the sphere. Relative orientation of adjacent images is determined by a linear transform and directions of all the images relative to the sphere are determined by multiplying the linear transform and directions of all the images relative to the sphere are determined by multiplying the linear transform matrix. Images are dynamically composed and projected back at a selected viewing direction without using an environment map. We have derived mathematical formula of the transform matrix. To improve the back projection speed, we also make environment maps by projecting images from the sphere to equator plane by setting projection center at southern or northern pole of the sphere. 3D scene can be projected to two circles in almost uniform pixel density without singularity at the poles. Distortion of lens causes large accumulated registration error. We developed a method for calibrating and correcting the distortion using matched corresponding points of adjacent images. A pyramid-based image matching method is also developed to reduce accumulated registration error.