Improving quality of autofluorescence images using non-rigid image registration

This work concerns quality improvement of auto-fluorescence retinal images by averaging of non-rigidly registered images. The necessity of using the elastic spatial transformation model is documented as well as the need for similarity criterion capable of dealing with the non-homogenous and variable illumination of retinal images. The presented multilevel registration algorithm provides parameters of primarily affine and then B-spline free-form spatial transformation optimal with respect to the mutual information similarity criterion. The registration was tested on three modeled image sets of 100 images. The difference of artificially introduced pre-deformation displacement field and the displacement field found by our algorithm clearly showed the ability to compensate for the diverse modeled distortions. Further, the registration algorithm was used for improving quality of realistic retinal images using averaging of registered frames of image sequences. The whole method was verified by processing of 16 time series of real images. The gain in signal to noise ratio in the averaged registered images with respect to individual frame reach the expected about 4dB, without introducing a visible blur. The final image was substantially less blurred than the non-registered averaged image, which is documented by comparison of the autocorrelation functions of both images.