Kernel Modeling Super-Resolution on Real Low-Resolution Images

Ruofan Zhou     Sabine Süsstrunk
Image and Visual Representation Lab
École Polytechnique Fédérale de Lausanne

- Supplementary Material -

project page

In Section 4.6 in the paper, we present our experiment with different zoom. Here we provide more details about the experiment.
We use a Nikon AF-S 24-70mm zoom lens to collect 3 pairs of images. The RGB image taken with a 70mm focal length serves as the 2X zoom ground truth of the raw sensor data taken with a 35 mm focal length. We set ISO equals to 400. We capture images with a distance of at least 100 meters to avoid perspective shifts.
A slight misalignment is unavoidable because of focal length variations in the center of projection when the lens zooms in and out. We align the bicubic upscaled 35mm "low-resolution" image with the "zoomed-in" groundtruth 70mm image by applying a grid search in horizontal and vertical shifts (within 100 pixels) as well as a stretching (range between 0.9 to 1.1).

Here we show three examples of super-resolution results:

SRCNN[3]	VDSR[4]	EDSR[1]	DBPN[2]	KMSR	refernece

SRCNN[3]	VDSR[4]	EDSR[1]	DBPN[2]	KMSR	refernece

SRCNN[3]	VDSR[4]	EDSR[1]	DBPN[2]	KMSR	refernece

Reference:
[1] Lim Bee, Son Sanghyun, Kim Heewon, Nah Seungjun. Lee Kyoung Mu. "Enhanced Deep Residual Networks for Single Image Super-Resolution", in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017
[2] Haris Muhammad, Shakhnarovich Greg, Ukita Norimichi. "Deep Back-Projection Networks for Super-Resolution", in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018
[3] Chao Dong, Chen Change Loy, Kaiming He, Xiaoou Tang. "Image Super-Resolution Using Deep Convolutional Networks", in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2015
[4] Jiwon Kim, Jung Kwon Lee, Kyoung Mu Lee. "Accurate Image Super-Resolution Using Very Deep Convolutional Networks", in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016