Towards visually convincing image segmentation

摘要

The aim of this work is to bring the computer marking of edges (on grey images) into closer correspondence with the perception of edges when an image is viewed on a typical computer screen under normal lighting conditions. To obtain the required correspondence, the computer should start off with comparable source data to the human, and perform similar edge-detecting operations on that data. Calculation of the amount of stimulation received by the eye requires knowledge of the screen's photopic emission, the amount of ambient light reflected from the screen, and the eye's sensitivity to luminance contrast. The contrast calculation used was derived from Michelson contrast by separating out a term for the veiling light, Lveil. The contrast, here called veiled-Michelson contrast CvM, is defined as CvM = (Ldmax − Ldmin/Ldmax + Ldmin + 2Lveil), where Ldmaxis the greater of the two photopic luminances emitted by the screen in darkness. The results of this calculation are compared with those of a step-difference operator, and show a qualitively improved correspondence with the visual perception.