Model-Based Illumination Correction

 M. Foracchia, E. Grisan, A. Ruggeri. Luminosity and contrast normalization in retinal images. Med Image Anal 9/3:179-190, 2005.

Retinal images are routinely acquired and assessed to provide diagnostic evidence for many important diseases. Because of the acquisition process, very often these images are non-uniformly illuminated and exhibit local luminosity and contrast variability. This problem may seriously affect the diagnostic process and its outcome, especially if an automatic computer-based procedure is used.

We propose here a new method to estimate and correct luminosity variation in retinal images. The method uses the hue, saturation, value (HSV) colour space to better decouple the luminance and chromatic information. Then, it fits an illumination model on a proper subregion (the retinal background) of the saturation and value channels. This solves many of the drawbacks of previously proposed methods, as filter-based correction which fails when large lesions or retinal features are present.

In retinal images two effects are usually present: We describe these effects by means of two elliptic paraboloids, whose parameters are estimated on different region of the fundus: the luminosity decrease in the interior part of the image, and the glare in the periphery. For physical reasons, we want the combination of the two models to be continuous and to have a smooth transition between the regions influenced by the two different illumination phenomena.

Paraboloid model for the internal region

Paraboloid model for the peripheral region

Sigmoidal function to ensure smoothness on the combination of the 2 previous paraboloids

Complete illumination model

From left to right, the original image, the estimated background pixels, and the estimated luminosity model for the value channel

Retinal images with the luminosity corrected with the proposed method, with mean and standard deviation imposed to be equal to those of the original images.

return to top