Extraction of Line Drawings from Gray Value Images by Non-Local Analysis of Edge Element Structures

Abstract

Edge elements defined as maxima of the gradient magnitude in gradient direction of a Gaussian-smoothed image are usually thresholded to suppress edge elements due to noise. In low contrast image regions, thresholding may suppress also edge elements which are part of a significant image structure and may thus result in the fragmentation or total loss of such structures. Based on an exhaustive categorization of edge element configurations in 5×5 pixel environments (about 34 million cases), straight and curved line structures are enhanced in edge element pictures without applying a uniform gradient magnitude thresholding. In contrast to the traditional pixel oriented gradient magnitude thresholding approach, we check for chaining of edge elements based mainly on the gradient direction. Using second moments of the change rates of gradient properties, we either reject edge element chains as noisy or accept them as a structure underlying the original image. The approach reported here has been designed deliberately for real-time execution. Experience with this approach applied to real world images demonstrates a significant improvement compared to a uniform thresholding approach.