Learning to Push the Limits of Efficient FFT-based Image Deconvolution - Supplemental Material

TitleLearning to Push the Limits of Efficient FFT-based Image Deconvolution - Supplemental Material
Publication TypeTechreport
Year of Publication2017
AuthorsKruse, J, Rother, C, Schmidt, U, Dresden, TU
  1. Details about boundary adjustment comparison Section 4.3 of the main paper compares our proposed boundary adjustment (BA) strategy (Our BA, cf. Eq. 17 and Fig. 2 of the main paper) to the traditional edgetapering method (ET once, cf. Eq. 11 of the main paper) and the BA approach (ET each) of CSF [3]; these BA strategies are compared within our FDN model, the CSF model, and a standard Wiener filter [5]. Specifically, we use the publicly available code to train different variants of the CSF model on a dataset of the same size as ours, and only adjust the BA strategy. Furthermore, we apply the Wiener filter as defined in Eq. 2 of the main paper, which we can use iteratively with our BA approach by replacing y with ϕ t (y, k, x t); we estimate the expected image spectrum n from 3000 clean image patches. While our BA comparison is depicted visually in Fig. 5 of the main paper, Table 1 also provides the numeric results and additionally includes stages 6-10 of our FDN model. As compared to the BA approach of CSF (ET each), the results suggest that CSF would also benefit from further stages if used with our BA strategy (cf. 6 th column). Remarkably, the performance of the Wiener filter is not even fully saturated after 50 iterations (Wiener 50) when applied with our BA approach (cf. 3 rd column, only every 5 th step shown after iteration 10). Fig. 1 shows an example where our proposed BA strategy yields a substantial improvement in image quality compared to standard edgetapering (ET once).
Citation KeyKruse