At&t Bell Laboratories

Timothy J Thompson

Yann Andre LeCun

Ofer Matan

William D Satterfield

Segmentation of characters in a character set (10), made by placing a dark mark against a light background (12), is accomplished by establishing a vertical pixel projection for each pixel column in the image. The vertical pixel projections are filtered with a decay parameter so those pixel columns which contain only background have the highest projection. Thereafter, a set of 'cut-points' (points of image segmentation) is obtained so that each cut-point coincides with a pixel column whose vertical pixel projection is both a local maxima and exceeds a predetermined threshold. The number of such cut-points is counted and if the number is not significantly greater than a predetermined number, the image is segmented along the cut-points. Otherwise, the vertical projections of those pixel columns coincident with the cut-points are filtered with a decreasing threshold to reduce the number of potential cut-points.

Method and apparatus for image segmentation

A flexible technique that is robust in the presence of noise and distortion is disclosed for locating and utilizing fiducial markings to locate predetermined areas within an image. Initially, a set of candidate markings is determined, one marking from within each of a plurality of predesignated areas of a scanned image. Thereafter, an estimate of a set of image distortion factors relative to an ideally registered image of the same type as the scanned image is derived. Predicted locations for the fiducial markings are derived using this estimated set of image distortion factors. If the difference in location between the actual location of any candidate markings and its predicted location is greater than a predetermined threshold, at least one such candidate is removed from the set of candidate markings. The estimation and removal processes are repeated, utilizing the new current set of candidate fiducial markings until all remaining candidate markings are within a predetermined distance of their predicted locations. Predetermined areas within the scanned image can thereafter be accurately located by combining the known location of the area within an ideal image with the final estimated set of image distortion factors. The predicted location may be further improved by using a weighted average of the ideal versus the actual fiducial marking differences, to account for local non-linear variations.

Growing community of inventors

Howell, NJ, United States of America

Timothy J Thompson