Image subtraction can be used in a wide variety of applications e.g. motion detection, but interpretation of the resulting difference image can be problematic since the pixel values have no objective meaning. We have developed a novel non-parametric image subtraction routine, which gives a difference image where the pixel values represent the probability that the pairing of pixel values at that position in the original images was drawn from the bulk distribution for pixel values in the images. This distribution is obtained in a non-parametric way using a grey-level scattergram of the original images. Furthermore, the probability distribution in the difference image is honest (i.e. a value of 0.1 implies that 10% of the pixels in the image were more unlikely to be drawn from the bulk distribution).
The images above show an example of the non-parametric image subtraction algorithm. The first and second images show MRI scans of the brain taken with different echo train times, producing a global difference between the scans. A grey-level offset of around twice the standard deviation of the noise in the images has been added to a small circular region of the second image. The altered region is difficult to see in the original images, and would be obscured in the result of a simple subtraction by the global differences between the images. However, non-parametric image subtraction detects the offset region as being statistically significant compared to the global differences between the images, and so shows the altered region clearly. I anticipate that this algorithm will have applications in many areas of machine vision, including tracking the progress of pre-clinical MS through volumetric analysis of MS lesions.
You can read a more detailed description of the new technique in my papers for BMVC 2000, the BMVC 2000 special edition of Image and Vision Computing, and MIUA 2001. Better still, download TINA from here and try it out yourself.