whb@castle.ed.ac.uk (H Bruce) (04/03/91)
I am compressing a 4 frame sequence of grey scale images 1 second apart by sending the first frame then subsequent frame differences. This works fine but if the lighting changes then the grey scale values change enough to signal a pixel difference when one has not occured. This can be overcome by calculating the mean (m0) and standard deviation (sd0) of the first image and applying them to the "normalised" pixel values of the subsequent images. ie p = (p-m)/sd * sd0 + m0 However if on object moves and changes size between frames the mean and standard deviation can change substantially so that two subsequent "normalised" images still have different illumination levels. To overcome this I can only guess that a frame difference is necessary before the normalisation so that parts of the image that vary are not included in the statistical calculations. Is this a sensible approach or can anyone recommend and alternative algorithm to overcome this problem ? Thanks, Henry Bruce.
mrd@ecs.soton.ac.uk (Mark Dobie) (04/05/91)
In <9444@castle.ed.ac.uk> whb@castle.ed.ac.uk (H Bruce) writes: >I am compressing a 4 frame sequence of grey scale images 1 second apart by >sending the first frame then subsequent frame differences. >This works fine but if the lighting changes then the grey scale values >change enough to signal a pixel difference when one has not occured. >This can be overcome by calculating the mean (m0) and standard deviation (sd0) >of the first image and applying them to the "normalised" pixel values of the >subsequent images. >ie p = (p-m)/sd * sd0 + m0 >However if on object moves and changes size between frames the mean and >standard deviation can change substantially so that two subsequent >"normalised" images still have different illumination levels. Try reading "Illumination Independent Change Detection for Real World Image Sequences" by Kurt Skifstad and Ramesh Jain in Computer Vision, Graphics and Image Processing, vol 46, pp 387-399 (1989) They discuss several methods with the particular aim of being robust against changes in illumination of the scene. They describe the normalisation technique you use applied to regions in the images. They continue with a method derived from a graphics shading model, which proves the overall best performer. Mark. -- Mark Dobie M.Dobie@uk.ac.soton.ecs (JANET) University of Southampton M.Dobie@ecs.soton.ac.uk (Bitnet)