Towards Quality Assessment of Blind Deconvolution with Shift Compensation
Ghada Laribi, Martin Welk · 2022 26th International Conference on Pattern Recognition (ICPR) · 2022
Quantitative assessment of restoration quality in blind deconvolution is not straightforward. The intricacy comes from the opposite shifts occurring during image reconstruction in both the reconstructed image and point-spread function. State of the art procedures attempting alignment lack specificity and might actually induce interpolation-based errors, whereas alignment-free approaches disregard the possibility of shift variability. Other methods introduce a superresolution-based MSE/PSNR measure involving non-integer shifts. We propose a method to estimate non-integer displacements between the ground-truth image and the reconstructed image via the Fourier domain, which enables at the same time to incorporate interpolation-free shift compensation into the computation of MSE/PSNR measures. We tested our method on synthetically shifted images with additive noise as well as deconvolution results with known shift. Results indicate that distortions of error measures are reduced compared to interpolation-based shift compensation methods, getting closer to a fair assessment of restoration quality without wrongly favouring any blind deconvolution method over another. It was observed that PSNR measures corrected with estimated displacements are close to those corrected with real displacements.