Correction to: Deep Unpaired Blind Image Super-Resolution Using Self-supervised Learning and Exemplar Distillation

1 Correction to: International Journal of Computer Vision https://doi.org/10.1007/s11263-023-01957-w

In the PDF of this article, the figure citation information was missing and should have been given as below.

• 1. Figure 1(c): “IKC [9]” should be “IKC (Gu et al., 2019)”, Fig. 1(d): “KMSR [48]” should be “KMSR (Zhou and Süsstrunk, 2019)”; Fig. 1(e): “KGAN [1] + ZSSR [33]” should be “KGAN (Bell-Kligler et al., 2019) + ZSSR (Shocher et al., 2018)”, Fig. 1(f): “MZSR [34]” should be “MZSR (Soh et al., 2020)”.

• 2. Figure 4(d): “EDSR [23]” should be “EDSR (Lim et al., 2017)”, Fig. 4(e): “RCAN [46]” should be “RCAN (Zhang et al., 2018b)”, Fig. 4(f): “IKC [9]” should be “IKC (Gu et al., 2019)”, Fig. 4(g): “KMSR [48]” should be “KMSR (Zhou and Süsstrunk, 2019)”; Fig. 4(h): “ZSSR [33]” should be “ZSSR (Shocher et al., 2018)”; Fig. 4(i): “KGAN + ZSSR” should be “KGAN (Bell-Kligler et al., 2019) + ZSSR (Shocher et al., 2018)”; Fig. 4(j): “MZSR [34]” should be “MZSR (Soh et al., 2020)”.

• 3. Figure 5(c): “KMSR [48]” should be “KMSR (Zhou and Süsstrunk, 2019)”; Fig. 5(d): “ZSSR [33]” should be “ZSSR (Shocher et al., 2018)”; Fig. 5(e): “KGAN [1] + ZSSR [33]” should be “KGAN (Bell-Kligler et al., 2019) + ZSSR (Shocher et al., 2018)”; Fig. 5(f): “MZSR [34]” should be “MZSR (Soh et al., 2020)”.

• 4. Figure 6(c): “KMSR [48]” should be “KMSR (Zhou and Süsstrunk, 2019)”; Fig. 6(d): “ZSSR [33]” should be “ZSSR (Shocher et al., 2018)”; Fig. 6(e): “KGAN [1] + ZSSR [33]” should be “KGAN (Bell-Kligler et al., 2019) + ZSSR (Shocher et al., 2018)”; Fig. 6(f): “MZSR [34]” should be “MZSR (Soh et al., 2020)”.

Fig. 1

Super-resolution results ($\times$4) on a synthetic image.We develop an effective method to solve blind image super-resolution when the paired training data is not available. Our approach first estimates the blur kernel and intermediate high-resolution (HR) image from the lowresolution (LR) one in a self-supervised learning manner and then refines the intermediate HR images based on a variational model. Finally, we develop an exemplar distillation module to constrain the deep models for better LR-to-HR mapping estimation. Our method does not require the exact LR-HR paired training data and recovers better structural details than both supervised and self-supervised learning methods

Fig. 4

Visual comparisons for 4× SR on one example from the Urban100 dataset (Huang et al., 2015). The proposed method generates much better results with fine detailed structures

Fig. 5

Visual comparisons on a real example ($\times$4). The proposed method recovers much clearer images with better characters

Fig. 6

Another super-resolution results on a real example ($\times$4). The proposed method recovers much clearer images with better characters

The original article PDF has now been corrected.