Purchase this article with an account.
Y Yitzhaky, E Peli; Visual Feature Based Registration of Images from Different Retinal Eccentricities . Invest. Ophthalmol. Vis. Sci. 2002;43(13):2839.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: Saccadic eye movements form sequences of displaced (unregistered) images of a scene, from different retinal eccentricities; yet, the perceived scene is stable and aligned (registered). This has been explained by the use of two information sources: neural (efference copy and proprioceptive) and retinal. Though no evidence was shown that one of these sources is solely responsible for the perceived stability, retinal information is likely to be involved due to its higher spatial precision. We implemented and analyzed the possible use of retinal information from different eccentricities as a basis for such image registration. Methods: A multi-channel feature extraction technique based on a visual model (Peli 2002, Proc. IEEE in press) was extended to include dependency on retinal eccentricity. Using displaced images simulated as cortical representations at different eccentricities, visual information intensity (energy at the spatial frequency band) was quantitatively evaluated with respect to eccentricity. Based on that, image features were detected and used for the registration and as a basis for speeding computation. The peak in the similarity measure as a function of displacement identifies the displacement between the images. The sharpness of the peak is a measure of the robustness of the registration. Results: As anticipated, less image details pass through the visual channels as spatial frequency and eccentricity increase. Image registration was computed using the visual features extracted from different eccentricities. Based on the retinal information a substantial speed improvement was achieved. A sharper peak (between 2.5 to 5 times) in the similarity measure function was obtained when visual features were used instead of the actual images. Conclusion: Image registration can potentially be carried out using the visually extracted features. The intensity of the retinal information according to the visual model provides means for speeding the registration. The sharper peak may also be advantageous in a noisy registration process, but this benefit is reduced by the noisier nature of the visually detected features.
This PDF is available to Subscribers Only