Toward Real-World Single Image Super-Resolution: A New Benchmark and A New Model
TToward Real-World Single Image Super-Resolution:A New Benchmark and A New Model
Jianrui Cai , ∗ , Hui Zeng , ∗ , Hongwei Yong , Zisheng Cao , Lei Zhang , , § The Hong Kong Polytechnic University, DJI Co.,Ltd, DAMO Academy, Alibaba Group { csjcai, cshzeng, cshyong, cslzhang } @comp.polyu.edu.hk, [email protected] Abstract
Most of the existing learning-based single image super-resolution (SISR) methods are trained and evaluatedon simulated datasets, where the low-resolution (LR) images are generated by applying a simple and uniformdegradation (i.e., bicubic downsampling) to their high-resolution (HR) counterparts. However, the degradationsin real-world LR images are far more complicated. As a consequence, the SISR models trained on simulated databecome less effective when applied to practical scenarios. In this paper, we build a real-world super-resolution(RealSR) dataset where paired LR-HR images on the same scene are captured by adjusting the focal length of adigital camera. An image registration algorithm is developed to progressively align the image pairs at differentresolutions. Considering that the degradation kernels are naturally non-uniform in our dataset, we present aLaplacian pyramid based kernel prediction network (LP-KPN), which efficiently learns per-pixel kernels to recoverthe HR image. Our extensive experiments demonstrate that SISR models trained on our RealSR dataset deliverbetter visual quality with sharper edges and finer textures on real-world scenes than those trained on simulateddatasets. Though our RealSR dataset is built by using only two cameras (Canon 5D3 and Nikon D810), the trainedmodel generalizes well to other camera devices such as Sony a7II and mobile phones.
1. Introduction
Single image super-resolution (SISR) [15] aims to recover a high-resolution (HR) image from its low-resolution(LR) observation. SISR has been an active research topic for decades [38, 57, 45, 47, 5] because of its high practicalvalues in enhancing image details and textures. Since SISR is a severely ill-posed inverse problem, learning imageprior information from the HR and/or LR exemplar images [15, 13, 55, 19, 14, 7, 24, 56, 11, 20, 46, 41] playsan indispensable role in recovering the details from an LR input image. Benefitting from the rapid developmentof deep convolutional neural networks (CNNs) [28], recent years have witnessed an explosive spread of trainingCNN models to perform SISR, and the performance has been consistently improved by designing new CNNarchitectures [9, 50, 42, 23, 44, 30, 62, 61] and loss functions [22, 29, 40].Though significant advances have been made, most of the existing SISR methods are trained and evaluatedon simulated datasets which assume simple and uniform degradation ( i.e ., bicubic degradation). Unfortunately,SISR models trained on such simulated datasets are hard to generalize to practical applications since the authenticdegradations in real-world LR images are much more complex [54, 26]. Fig. 1 shows the SISR results of a real-world image captured by a Sony a7II camera. We utilize the state-of-the-art RCAN method [61] to train three SISR ∗ The first two authors contribute equally to this work. A part of this dataset was used in the image super-resolution challenge in theNTIRE 2019 challenges [1]. § Corresponding author: Lei Zhang a r X i v : . [ c s . C V ] A p r a) Image captured by Sony a7II (b) Bicubic(c) RCAN + BD(d) RCAN + MD (e) RCAN + RealSR (f) LP-KPN + RealSR Figure 1. The SISR results ( × ) of (a) a real-world image captured by a Sony a7II camera. SISR results generated by(b) bicubic interpolator, RCAN models [61] trained using image pairs (in DIV2K [45]) with (c) bicubic degradation (BD),(d) multiple simulated degradations (MD) [60], and (e) authentic distortions in our RealSR dataset. (f) SISR result by theproposed LP-KPN model trained on our dataset. Note that our RealSR dataset is collected by Canon 5D3 and Nikon D810cameras. models using simulated image pairs (in DIV2K [45]) with bicubic degradation, multiple simulated degradations[60] and image pairs with authentic distortions in our dataset to be constructed in this paper. The results clearlyshow that, compared with the simple bicubic interpolator (Fig. 1(b)), the RCAN models trained on simulateddatasets (Figs. 1(c) ∼ conv layers) can output more distinct result thanRCAN (over 400 conv layers) using much fewer layers.The contributions of this work are twofold: • We build a RealSR dataset consisting of HR and LR image pairs with different scaling factors. It provides,to the best of our knowledge, the first general purpose benchmark for real-world SISR model training andevaluation. • We present an LP-KPN model and validate its efficiency and effectiveness in real-world SISR.Extensive experiments are conducted to quantitatively and qualitatively analyze the performance of our RealSRdataset in training SISR models. Though the dataset in its current version is built using only two cameras, thetrained SISR models exhibit good generalization capability to images captured by other types of camera devices.
2. Related Work
SISR datasets.
There are several popular datasets, including Set5 [4], Set14 [59], BSD300 [32], Urban100 [19],Manga109 [33] and DIV2K [45] that have been widely used for training and evaluating the SISR methods. Inall these datasets, the LR images are generally synthesized by a simple and uniform degradation process such asbicubic downsampling or Gaussian blurring followed by direct downsampling [10]. The SISR Models trainedon these simulated data may exhibit poor performance when applied to real LR images where the degradationdeviates from the simulated ones [12]. To improve the generalization capability, Zhang et al . [60] trained theirmodel using multiple simulated degradations and Bulat et al . [6] used a GAN [17] to generate the degradationprocess. Although these more advanced methods can simulate more complex degradation, there is no guaranteethat such simulated degradation can approximate the authentic degradation in practical scenarios which is usuallyvery complicated [26].It is very challenging to get the ground-truth HR image for an LR image in real-world scenarios, making thetraining and evaluation of real-world SISR models difficult. To the best of our knowledge, only two recent attemptshave been made on capturing real-world image pairs for SISR. Qu et al . [39] put two cameras together with abeam splitter to collect a dataset with paired face images. K¨ohler et al . [26] employed hardware binning on thesensor to capture LR images and used multiple postprocessing steps to generate different versions of an LR image.However, both the datasets were collected in indoor laboratory environment and very limited number of scenes(31 face images in [39] and 14 scenes in [26]) were included. Different from them, our dataset is constructed byadjusting the focal length of DSLR cameras which naturally results in image pairs at different resolutions, and itcontains scenes in both indoor and outdoor environments.3 igure 2. Illustration of thin lens. u, v, f represent the object distance, image distance and focal length, respectively. h and h denote the size of object and image. Kernel prediction networks.
Considering that the degradation kernel in our RealSR dataset is spatially variant,we propose to train a kernel prediction network (KPN) for real-world SISR. The idea of KPN was first proposedin [2] to denoise Monte Carlo renderings and it has proven to have faster convergence and better stability thandirect prediction [48]. Mildenhall et al . [34] trained a KPN model for burst denoising and obtained state-of-the-artperformance on both synthetic and real data. Similar ideas have been employed in estimating the blur kernelsin dynamic deblurring [43, 16] or convolutional kernels in video interpolation [35, 36]. We are among the firstto train a KPN for SISR and we propose the LP-KPN to perform kernel prediction in the scale space with highefficiency.
3. Real-world SISR Dataset
To build a dataset for learning and evaluating real-world SISR models, we propose to collect images of the samescene by adjusting the lens of DSLR cameras. Sophisticated image registration operations are then performed togenerate the HR and LR pairs of the same content. The detailed dataset construction process is presented in thissection.
The DSLR camera imaging system can be approximated as a thin lens [53]. An illustration of the imageformation process by thin lens is shown in Fig. 2. We denote the object distance, image distance and focal lengthby u, v, f , and denote the size of object and image by h and h , respectively. The lens equation is defined asfollows [53]: f = 1 u + 1 v . (1)The magnification factor M is defined as the ratio of the image size to the object size: M = h h = vu . (2)In our case, the static images are taken at a distance ( i.e ., u ) larger than 3.0m. Both h and u are fixed and u ismuch larger than f (the largest f is 105mm). Combining Eq. (1) and Eq. (2), and considering the fact that u (cid:29) f ,we have: h = fu − f h ≈ fu h . (3)Therefore, h is approximately linear to f . By increasing the focal length f , larger images with finer details willbe recorded in the camera sensor. The scaling factor can also be controlled (in theory) by choosing specific valuesof f . 4 able 1. Number of image pairs for each camera at each scaling factor.Camera Canon 5D3 Nikon D810Scale × × × × × × We used two full frame DSLR cameras (Canon 5D3 and Nikon D810) to capture images for data collection. Theresolution of Canon 5D3 is × , and that of Nikon D810 is × . To cover the common scalingfactors ( e.g ., × , × , × ) used in most previous SISR datasets, both cameras were equipped with one 24 ∼ f /4.0 zoom lens. For each scene, we took photos using four focal lengths: 105mm, 50mm, 35mm, and 28mm.Images taken by the largest focal length are used to generate the ground-truth HR images, and images taken by theother three focal lengths are used to generate the LR versions. We choose 28mm rather than 24mm because lensdistortion at 24mm is more difficult to correct in post-processing, which results in less satisfied quality in imagepair registration.The camera was set to aperture priority mode and the aperture was adjusted according to the depth-of-field(DoF) [52]. Basically, the selected aperture value should make the DoF large enough to cover the scene and avoidsevere diffraction. Small ISO is preferred to alleviate noise. The focus, white balance, and exposure were setto automatic mode. The center-weighted metering option was selected since only the center region of capturedimages were used in our final dataset. For stabilization, the camera was fixed on a tripod and a bluetooth remotecontroller was used to control the shutter. Besides, lens stabilization was turned off and the reflector was pre-risedwhen taking photos.To ensure the generality of our dataset, we took photos in both indoor and outdoor environment. Scenes withabundant texture are preferred considering that the main purpose of super-resolution is to recover or enhance imagedetails. For each scene, we first captured the image at 105mm focal length and then manually decreased the focallength to take three down-scaled versions. scenes were captured, and there are no overlapped scenes betweenthe two cameras. After discarding images having moving objects, inappropriate exposure, and blur, we have HR and LR image pairs in total. The numbers of image pairs for each camera at each scaling factor are listed inTable 1.
Although it is easy to collect images on different scales by zooming the lens of a DSLR camera, it is diffi-cult to obtain pixel-wise aligned image pairs because the zooming of lens brings many uncontrollable changes.Specifically, images taken at different focal lengths suffer from different lens distortions and usually have differentexposures. Moreover, the optical center will also shift when zooming the focal length because of the inherentdefect of lens. Even the scaling factors are varying slightly because the lens equation (Eq. (1)) cannot be preciselysatisfied in practical focusing process. With the above factors, none of the existing image registration algorithmscan be directly used to obtain accurate pixel-wise registration of two images captured under different focal length.We thus develop an image registration algorithm to progressively align such image pairs to build our RealSRdataset.The registration process is illustrated in Fig. 3. We first import the images with meta information into PhotoShopto correct the lens distortion. However, this step cannot perfectly correct the lens distortion especially for the regiondistant from the optical center. We thus further crop the interested region around the center of the image, wheredistortion is not severe and can be well corrected. The cropped region from the image taken at 105mm focal lengthis used as the ground-truth HR image, whose LR counterparts are to be registered from the original images takenat 50mm, 35mm, or 28mm focal length. Since there is certain luminance and scale difference between imagestaken at different focal lengths, those popular keypoint based image registration algorithms such as SURF [3] and5 istortion correction & central region crop Iterative registrationImage taken at 105mmImage taken at 28mm Ground-truth HR image Aligned LR image
As reference
Distortion correction & central region crop LR image
Figure 3. Illustration of our image pair registration process.
SIFT [31] cannot always achieve pixel-wise registration, which is necessary for our dataset. To obtain accurateimage pair registration, we develop a pixel-wise registration algorithm which simultaneously considers luminanceadjustment. Denote by I H and I L the HR image and the LR image to be registered, our algorithm minimizes thefollowing objective function: min τ || αC ( τ ◦ I L ) + β − I H || pp , (4)where τ is an affine transformation matrix, C is a cropping operation which makes the transformed I L have thesame size as I H , α and β are luminance adjustment parameters, || · || p is a robust L p -norm ( p ≤ , e.g ., L -norm.The above objective function is solved in an iterative manner. At the beginning, according to Eq. (3), the τ is initialized as a scaling transformation with scaling factor calculated as the ratio of two focal lengths. Let I (cid:48) L = C ( τ ◦ I L ) . With I (cid:48) L and I H fixed, the parameters for luminance adjustment can be obtained by α = std ( I H ) / std ( I (cid:48) L ) and β = mean ( I H ) − α mean ( I (cid:48) L ) , which can ensure I (cid:48) L having the same pixel mean and varianceas I H after luminance adjustment. Then we solve the affine transformation matrix τ with α and β fixed. Ac-cording to [37, 58], the objective function w.r.t. τ is nonlinear, which can be iteratively solved by a locally linearapproximation: min ∆ τ || αC ( τ ◦ I L ) + β + α J ∆ τ − I H || pp , (5)where J is the Jacobian matrix of C ( τ ◦ I L ) w.r.t. τ , and this objective function can be solved by an iterativelyreweighted least square problem (IRLS) as follows [8]: min ∆ τ || w (cid:12) ( A ∆ τ − b ) || , (6)where A = α J , b = I H − ( αC ( τ ◦ I L ) + β ) , w is the weight matrix and (cid:12) denotes element-wise multiplication.Then we can obtain: ∆ τ = ( A (cid:48) diag ( w ) A ) − A (cid:48) diag ( w ) b , (7)and τ can be updated by: τ = τ + ∆ τ .We iteratively estimate the luminance adjustment parameters and the affine transformation matrix. The opti-mization process converges within iterations since our prior information of the scaling factor provides a goodinitialization of τ . After convergence, we can obtain the aligned LR image as I AL = αC ( τ ◦ I L ) + β .
4. Laplacian Pyramid based Kernel Prediction Network
In Section 3, we have constructed a new real-world super-resolution (RealSR) dataset, which consists of pixel-wise aligned HR and LR image pairs { I H , I AL } of size h × w . Now the problem turns to how to learn an effectivenetwork to enhance I AL to I H . For LR images in our RealSR dataset, the blur kernel varies with the depth in a6 igure 4. Framework of the Laplacian pyramid based kernel prediction network. By decomposing the image into a Laplacianpyramid, using small kernels can leverage rich neighborhood information for super-resolution. scene [51] and the DoF [52] changes with the focal length. Training an SISR model which directly transformsthe LR image to the HR image, as done in most of the previous CNN based SISR methods, may not be thecost-effective way. We therefore propose to train a kernel prediction network (KPN) which explicitly learns anindividual kernel for each pixel. Compared with those direct pixel synthesis networks, KPN has proven to haveadvantages in efficiency, interpretability and generalization capability in tasks of denoising, dynamic deblurring,etc., [2, 34, 48, 43, 16, 27].The KPN takes the I AL as input and outputs a kernel tensor T ∈ R ( k × k ) × h × w , in which each vector in channeldimension T ( i, j ) ∈ R ( k × k ) can be reshaped into a k × k kernel K ( i, j ) . The reshaped per-pixel kernel K ( i, j ) isapplied to the k × k neighborhood of each pixel in the input LR image I AL ( i, j ) to reproduce the HR output. Thepredicted HR image, denoted by I PH , is obtained by: I PH ( i, j ) = (cid:104) K ( i, j ) , V ( I AL ( i, j )) (cid:105) , (8)where V ( I AL ( i, j )) represents a k × k neighborhood of pixel I AL ( i, j ) and (cid:104) · (cid:105) denotes the inner product operation.Eq. (8) shows that the output pixel is a weighted linear combination of the neighboring pixels in the inputimage. To obtain good performance, a large kernel size is necessary to leverage richer neighborhood information,especially when only a single frame image is used. On the other hand, the predicted kernel tensor T grows quadrat-ically with the kernel size k , which can result in high computational and memory cost in practical applications. Inorder to train a both effective and efficient KPN, we propose a Laplacian pyramid based KPN (LP-KPN).The framework of our LP-KPN is shown in Fig. 4. As in many SR methods [30, 47], our model works onthe Y channel of YCbCr space. The Laplacian pyramid decomposes an image into several levels of sub-imageswith downsampled resolution and the decomposed images can exactly reconstruct the original image. Using thisproperty, the Y channel of an LR input image I AL is decomposed into a three-level image pyramid { S , S , S } ,where S ∈ R h × w , S ∈ R h × w , and S ∈ R h × w . Our LP-KPN takes the LR image as input and predictsthree kernel tensors { T , T , T } for the image pyramid, where T ∈ R ( k × k ) × h × w , T ∈ R ( k × k ) × h × w , and T ∈ R ( k × k ) × h × w . The learned kernel tensors { T , T , T } are applied to the corresponding image pyramid7 S , S , S } , using the operation in Eq. (8), to restore the Laplacian decomposition of HR image at each level.Finally, the Laplacian pyramid reconstruction is conducted to obtain the HR image. Benefitting from the Laplacianpyramid, learning three k × k kernels can equally lead to a receptive field with size k × k at the original resolution,which significantly reduces the computational cost compared to directly learning one k × k kernel.The backbone of our LP-KPN consists of residual blocks, with each residual block containing convolu-tional layers and a ReLU function (similar structure to [30]). To improve the efficiency, we shuffle [42] the inputLR image with factor (namely, the h × w image is shuffled to h × w images) and input the shuffled images tothe network. Most convolutional blocks are shared by the three levels of kernels except for the last few layers. One × and one × shuffle operation are performed to upsample the spatial resolution of the latent image representa-tions at two lower levels, followed by individual convolutional blocks. Our LP-KPN has a total of convolutionallayers, which is much less than the previous state-of-the-art SISR models [30, 62, 61]. The detailed network ar-chitecture can be found in the supplementary material . The L -norm loss function L ( I H , I PH ) = || I H − I PH || isemployed to minimize the pixel-wise distance between the model prediction I PH and the ground-truth HR image I H .
5. Experiments
Experimental setup.
The number of image pairs in our RealSR dataset is reported in Table 1. We randomlyselected image pairs at each scaling factor for each camera to form the testing set, while using the remainingimage pairs as training set. Except for cross-camera testing, images from both the Canon and Nikon cameras werecombined for training and testing. Following the previous work [30, 61, 47], the SISR results were evaluated usingPSNR and SSIM [49] indices on the Y channel in the YCbCr space. The height and width of images lie in therange of [700, 3100] and [600, 3500], respectively. We cropped the training images into × patches to trainall the models. Data augmentation was performed by randomly rotating ◦ , ◦ , ◦ and horizontally flippingthe input. The mini-batch size in all the experiments was set to .All SISR models were initialized using the method in [18]. The Adam solver [25] with the default parameters( β = 0 . , β = 0 . and (cid:15) = 10 − ) was adopted to optimize the network parameters. The learning rate wasfixed at − and all the networks were trained for , K iterations. All the comparing models were trainedusing the Caffe [21] toolbox, and tested using Caffe MATLAB interface. All the experiments were conducted ona PC equipped with an Intel Core i7-7820X CPU, 128G RAM and a single Nvidia Quadro GV100 GPU (32G). To demonstrate the advantages of our RealSR dataset, we conduct experiments to compare the real-world super-resolution performance of SISR models trained on simulated datasets and RealSR dataset. Considering that moststate-of-the-art SISR models were trained on DIV2K [45] dataset, we employed the DIV2K to generate simulatedimage pairs with bicubic degradation (BD) and multiple degradations (MD) [60]. We selected three representativeand state-of-the-art SISR networks, i.e ., VDSR [23], SRResNet [29] and RCAN [61], and trained them on the BD,MD and RealSR training datasets for each of the three scaling factors ( × , × , × ), leading to a total of SISRmodels. To keep the network structures of SRResNet and RCAN unchanged, the input images were shuffled withfactor , , for the three scaling factors × , × , × , respectively.We applied the trained SISR models to the RealSR testing set, and the average PSNR and SSIM indicesare listed in Table 2. The baseline bicubic interpolator is also included for comparison. One can see that, on ourRealSR testing set, the VDSR and SRResNet models trained on the simulated BD dataset can only achieve com-parable performance to the simple bicubic interpolator. Training on the MD dataset brings marginal improvementsover BD, which indicates that the authentic degradation in real-world images is difficult to simulate. Employinga deeper architecture, the RCAN ( >
400 layers) can improve (0.2dB ∼ able 2. Average PSNR (dB) and SSIM indices on our RealSR testing set by different methods (trained on different datasets). Metric Scale Bicubic VDSR [23] SRResNet [29] RCAN [61]BD MD Our BD MD Our BD MD OurPSNR × .
61 32 .
63 32 .
65 33 .
64 32 .
66 32 .
69 33 .
69 32 .
91 32 .
92 33 . × .
34 29 .
40 29 .
43 30 .
14 29 .
46 29 .
47 30 .
18 29 .
66 29 .
69 30 . × .
99 28 .
03 28 .
06 28 .
63 28 .
09 28 .
12 28 .
67 28 .
28 28 .
31 28 . SSIM × .
907 0 .
907 0 .
908 0 .
917 0 .
908 0 .
909 0 .
919 0 .
910 0 .
912 0 . × .
841 0 .
842 0 .
845 0 .
856 0 .
844 0 .
846 0 .
859 0 .
847 0 .
851 0 . × .
806 0 .
806 0 .
807 0 .
821 0 .
806 0 .
808 0 .
824 0 .
811 0 .
813 0 . Image captured by Canon 5D3 HR Bicubic SRResNet + BD SRResNet + MD SRResNet + RealSRVDSR + MD VDSR + RealSR RCAN + BD RCAN + MD RCAN + RealSRImage captured by Nikon D810 HR Bicubic SRResNet + BD SRResNet + MD SRResNet + RealSRVDSR + MD VDSR + RealSR RCAN + BD RCAN + MD RCAN + RealSR
Figure 5. SR results ( × ) on our RealSR testing set by different methods (trained on different datasets). performance than those trained on BD and MD datasets for all the three scaling factors. Specifically, for scalingfactor × , the models trained on our RealSR dataset have about . dB improvement on average for all the threenetwork architectures. The advantage is also significant for scaling factors × and × . In Fig. 5, we visualize thesuper-resolved images obtained by different models. As can be seen, the SISR results generated by models trainedon simulated BD and MD datasets tend to have blurring edges with obvious artifacts. On the contrary, modelstrained on our RealSR dataset recover clearer and more natural image details. More visual examples can be foundin the supplementary file . To demonstrate the efficiency and effectiveness of the proposed LP-KPN, we then compare it with SISRmodels, including VDSR, SRResNet, RCAN, a baseline direct pixel synthesis (DPS) network and four KPNmodels with kernel size k = 5 , , , . The DPS and the four KPN models share the same backbone as our LP-9 able 3. Average PSNR (dB) and SSIM indices for different models (trained on our RealSR training set) on our RealSRtesting set. Method PSNR SSIM × × × × × × Bicubic .
61 29 .
34 27 .
99 0 .
907 0 .
841 0 . VDSR [23] .
64 30 .
14 28 .
63 0 .
917 0 .
856 0 . SRResNet [29] .
69 30 .
18 28 .
67 0 .
919 0 .
859 0 . RCAN [61] .
87 30 .
40 28 .
88 0 .
922 0 .
862 0 . DPS .
71 30 .
20 28 .
69 0 .
919 0 .
859 0 . KPN, k = .
75 30 .
26 28 .
74 0 .
920 0 .
860 0 . KPN, k = .
78 30 .
29 28 .
78 0 .
921 0 .
861 0 . KPN, k =
13 33 .
83 30 .
35 28 .
85 0 .
923 0 .
862 0 . KPN, k =
19 33 .
86 30 .
39 28 .
90 0 .
924 0 .
864 0 . Our, k = .
90 30 .
42 28 .
92 0 .
927 0 .
868 0 . KPN. All models are trained and tested on our RealSR dataset. The PSNR and SSIM indices of all the competingmodels as well as the bicubic baseline are listed in Table 3.One can notice that among the four direct pixel synthesis networks ( i.e ., VDSR, SRResNet, RCAN and DPS),RCAN obtains the best performance because of its very deep architecture (over 400 layers). Using the samebackbone with less than layers, the KPN with × kernel size already outperforms the DPS. Using largerkernel size consistently brings better results for the KPN architecture, and it obtains comparable performance tothe RCAN when the kernel size increases to . Benefitting from the Laplacian pyramid decomposition strategy,our LP-KPN using three different × kernels achieves even better results than the KPN with × kernel. Theproposed LP-KPN obtains the best performance but with the lowest computational cost for all the three scalingfactors. The detailed complexity analysis and visual examples of the SISR results by the competing models canbe found in the supplementary file . To evaluate the generalization capability of SISR models trained on our RealSR dataset, we conduct a cross-camera testing. Images taken by two cameras are divided into training and testing sets, separately, with testingimages for each camera at each scaling factor. The three scales of images are combined for training, and modelstrained on one camera are tested on the testing sets of both cameras. The LP-KPN and RCAN models are comparedin this evaluation, and the PSNR indexes are reported in Table 4.It can be seen that for both RCAN and LP-KPN, the cross-camera testing results are comparable to the in-camera setting with only about . dB and . dB gap, respectively, while both are much better than bicubicinterpolator. This indicates that the SISR models trained on one camera can generalize well to the other camera.This is possibly because our RealSR dataset contains various degradations produced by the camera lens and imageformation process, which share similar properties across cameras. Between RCAN and LP-KPN models, theformer has more parameters and thus is easier to overfit to the training set, delivering slightly worse generalizationcapability than LP-KPN. Similar observation has been found in [2, 48, 34]. To further validate the generalization capability of our RealSR dataset and LP-KPN model, we evaluate ourtrained model as well as several competitors on images outside our dataset, including images taken by one Sonya7II DSLR camera and two mobile cameras ( i.e ., iPhone X and Google Pixel 2). Since there are no ground-truthHR versions of these images, we visualize the super-resolved results in Fig. 1 and Fig. 6. In all these cases, the10 able 4. Average PSNR (dB) index for cross-camera evaluation.
Tested Scale Bicubic RCAN LP-KPN(Trained) (Trained)Canon Nikon Canon NikonCanon × .
05 34 .
34 34 .
11 34 .
38 34 . × .
67 30 .
65 30 .
28 30 .
69 30 . × .
31 29 .
46 29 .
04 29 .
48 29 . Nikon × .
66 32 .
01 32 .
30 32 .
05 32 . × .
63 29 .
30 29 .
75 29 .
34 29 . × .
28 27 .
98 28 .
12 28 .
01 28 . Image captured by iPhone X Bicubic RCAN + BD RCAN + MDRCAN + RealSR KPN ( k = 19) + RealSR LP-KPN + RealSRImage captured by Google Pixel 2 Bicubic RCAN + BD RCAN + MDRCAN + RealSR KPN ( k = 19) + RealSR LP-KPN + RealSR Figure 6. SISR results ( × ) of real-world images outside our dataset. Images are captured by iPhone X and Google Pixel 2. LP-KPN trained on our RealSR dataset obtains better visual quality than the competitors, recovering more naturaland clearer details. More examples can be found in the supplementary file .11 . Conclusion
It has been a long standing problem for SISR research that the models trained on simulated datasets can hardlybe generalized to real-world images. We made a good attempt to address this issue, and constructed a real-world super-solution (RealSR) dataset with authentic degradations. One Canon and one Nikon cameras were usedto collect 595 HR and LR image pairs, and an effective image registration algorithm was developed to ensureaccurate pixel-wise alignment between image pairs. A Laplacian pyramid based kernel prediction network wasalso proposed to perform efficient and effective real-world SISR. Our extensive experiments validated that themodels trained on our RealSR dataset can lead to much better real-world SISR results than trained on existingsimulated datasets, and they have good generalization capability to other cameras. In the future, we will enlargethe RealSR dataset by collecting more image pairs with more types of cameras, and investigate new SISR modeltraining strategies on it.
7. Supplementary Material
Currently, the proposed RealSR dataset contains
HR-LR image pairs covering a variety of image contents.To ensure the diversity of our RealSR dataset, images are captured in indoor, outdoor and laboratory environments.Several examples of our RealSR dataset are shown in Fig. 7. It provides, to the best of our knowledge, the firstgeneral purpose benchmark for real-world SISR model training and evaluation. The RealSR dataset will be madepublicly available. (a) Laboratory scenes(b) Indoor scenes(c) Outdoor scenes
Figure 7. Some sample images of our RealSR dataset. .2. The details of our network architecture The network architecture of our proposed Laplacian pyramid based kernel prediction network (LP-KPN) isshown in Table 5. In this table, “ H × W × C conv” denotes a convolutional layer with C filters of size H × W which is immediately followed by a ReLU nonlinearity. Each residual block contains two × convolutionallayers with the same number of filters on both layers. The stride size for all convolution layers is set to and thenumber of filters C in each layer is set to , except for the last layer where C is set to . The structure of theresidual block is shown in Fig. 8, which is same as [30]. We use the shuffle operation [42] to downsample andupsample the the image. Table 5. Network architecture of the proposed LP-KPN.
Layer Activation size
Input × Shuffle, /4 × × × × conv, pad 1 × × × Residual blocks, 64 filters × × × × conv, pad 1 × × Shuffle, × Shuffle, × - × × × ×
16 48 × × × × conv, pad 1 × × conv, pad 1 × × conv, pad 1 × ×
64 96 × ×
64 48 × × × × conv, pad 1 × × conv, pad 1 × × conv, pad 1 × ×
64 96 × ×
64 48 × × × × conv, pad 1 × × conv, pad 1 × × conv, pad 1 × ×
25 96 × ×
25 48 × × Per-pixel conv by Eq. (8) Per-pixel conv by Eq. (8) Per-pixel conv by Eq. (8) ×
192 96 ×
96 48 × Output (Laplacian pyramid reconstruction) × Figure 8. Residual block used in our network.
In this subsection, we provide more visual results by SISR models trained on simulated SISR datasets (BDand MD [60]) and our proposed RealSR dataset. Two images captured by Canon 5D3, two images captured byNikon D810 and their super-resolved results are shown in Fig. 9. Again, the models trained on our RealSR datasetconsistently obtain better visual quality compared to their counterparts trained on simulated datasets.
The running time and the number of parameters of the competing models are listed in Table 6. One can see thatalthough larger kernel size can consistently bring better results for the KPN architecture, the number of parameterswill also greatly increase. Benefitting from the Laplacian pyramid decomposition strategy, our LP-KPN using × kernel can achieve better results than the KPN using × kernel, and it uses much less parameters. Specifically,our LP-KPN model contains less than parameters of the RCAN model [61] and it runs about 3 times faster thanRCAN. The visual examples of the SISR results by the competing models are shown in Fig. 10. Though all the13 mage captured by Canon 5D3 HR Bicubic SRResNet + BD SRResNet + MD SRResNet + RealSRVDSR + MD VDSR + RealSR RCAN + BD RCAN + MD RCAN + RealSRImage captured by Canon 5D3 HR Bicubic SRResNet + BD SRResNet + MD SRResNet + RealSRVDSR + MD VDSR + RealSR RCAN + BD RCAN + MD RCAN + RealSRImage captured by Nikon D810 HR Bicubic SRResNet + BD SRResNet + MD SRResNet + RealSRVDSR + MD VDSR + RealSR RCAN + BD RCAN + MD RCAN + RealSRImage captured by Nikon D810 HR Bicubic SRResNet + BD SRResNet + MD SRResNet + RealSRVDSR + MD VDSR + RealSR RCAN + BD RCAN + MD RCAN + RealSR Figure 9. SR results ( × ) on our RealSR testing set by different methods (trained on different datasets). Table 6. PSNR, SSIM, running time and parameters for different models (trained on our RealSR training set) on our RealSRtesting set. The running time is measured for an image of size × . We use the file size of Caffe model to representthe number of parameters. VDSR [23] SRResNet [29] RCAN [61] DPS KPN, k = KPN, k = KPN, k = KPN, k = Our, k = PSNR × .
64 33 .
69 33 .
87 33 .
71 33 .
75 33 .
78 33 .
83 33 .
86 33 . × .
14 30 .
18 30 .
40 30 .
20 30 .
26 30 .
29 30 .
35 30 .
39 30 . × .
63 28 .
67 28 .
88 28 .
69 28 .
74 28 .
78 28 .
85 28 .
90 28 . SSIM × .
917 0 .
919 0 .
922 0 .
919 0 .
920 0 .
921 0 .
923 0 .
924 0 . × .
856 0 .
859 0 .
862 0 .
859 0 .
860 0 .
861 0 .
862 0 .
864 0 . × .
821 0 .
824 0 .
826 0 .
824 0 .
826 0 .
827 0 .
828 0 .
830 0 . Parameters . M . M . M . M . M . M . M . M . MTimes (sec.) . . . . . . . . . Image captured by Canon 5D3 HR Bicubic VDSR SRResNetRCAN Direct Synth KPN ( k = 19 ) OurImage captured by Nikon D810 HR Bicubic VDSR SRResNetRCAN Direct Synth KPN ( k = 19 ) Our Figure 10. SR results ( × ) on our RealSR testing set by different methods (all trained on our RealSR dataset). It can beseen that all SISR models trained on our RealSR dataset achieve good results, while our LP-KPN still obtains the best visualquality. mage captured by Sony a7II Bicubic RCAN + BD RCAN + MDRCAN + RealSR KPN ( k = 19) + RealSR LP-KPN + RealSRImage captured by iPhone X Bicubic RCAN + BD RCAN + MDRCAN + RealSR KPN ( k = 19) + RealSR LP-KPN + RealSRImage captured by Google Pixel 2 Bicubic RCAN + BD RCAN + MDRCAN + RealSR KPN ( k = 19) + RealSR LP-KPN + RealSR Figure 11. SISR results ( × ) of real-world images outside our dataset. Images are captured by Sony a7II, iPhone X andGoogle Pixel 2. .5. More super-resolved results on images outside our dataset In this subsection, we provide more super-resolved results on images outside our dataset, including imagestaken by one Sony a7II DSLR camera and two mobile cameras ( i.e ., iPhone X and Google Pixel 2). The visualexamples are shown in Fig. 11.
References [1] . 1[2] S. Bako, T. Vogels, B. McWilliams, M. Meyer, J. Nov´ak, A. Harvill, P. Sen, T. Derose, and F. Rous-selle. Kernel-predicting convolutional networks for denoising monte carlo renderings.
ACM Trans. Graph. ,36(4):97–1, 2017. 3, 4, 7, 10[3] H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust features. In
ECCV , pages 404–417, 2006.5[4] M. Bevilacqua, A. Roumy, C. Guillemot, and M. L. Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 3[5] Y. Blau, R. Mechrez, R. Timofte, T. Michaeli, and L. Zelnik-Manor. The 2018 PIRM challenge on perceptualimage super-resolution. In
ECCV , 2018. 1[6] A. Bulat, J. Yang, and G. Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do imagedegradation first. In
ECCV , 2018. 3[7] H. Chang, D.-Y. Yeung, and Y. Xiong. Super-resolution through neighbor embedding. In
CVPR , 2004. 1[8] R. Chartrand and W. Yin. Iteratively reweighted algorithms for compressive sensing. In
ICASSP , 2008. 6[9] C. Dong, C. C. Loy, K. He, and X. Tang. Learning a deep convolutional network for image super-resolution.In
ECCV , 2014. 1[10] W. Dong, L. Zhang, G. Shi, and X. Li. Nonlocally centralized sparse representation for image restoration.
IEEE Transactions on Image Processing , 22(4):1620–1630, 2013. 3[11] W. Dong, L. Zhang, G. Shi, and X. Wu. Image deblurring and super-resolution by adaptive sparse domainselection and adaptive regularization.
IEEE Transactions on Image Processing , 20(7):1838–1857, 2011. 1[12] N. Efrat, D. Glasner, A. Apartsin, B. Nadler, and A. Levin. Accurate blur models vs. image priors in singleimage super-resolution. In
ICCV , 2013. 3[13] G. Freedman and R. Fattal. Image and video upscaling from local self-examples.
ACM Transactions onGraphics (TOG) , 30(2):12, 2011. 1[14] W. T. Freeman, E. C. Pasztor, and O. T. Carmichael. Learning low-level vision.
International journal ofcomputer vision , 40(1):25–47, 2000. 1[15] D. Glasner, S. Bagon, and M. Irani. Super-resolution from a single image. In
ICCV , pages 349–356, 2009. 1[16] D. Gong, J. Yang, L. Liu, Y. Zhang, I. Reid, C. Shen, A. Van Den Hengel, and Q. Shi. From motion blur tomotion flow: a deep learning solution for removing heterogeneous motion blur. In
CVPR , 2017. 3, 4, 7[17] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio.Generative adversarial nets. In
NIPs , 2014. 3[18] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance onimagenet classification. In
ICCV , 2015. 8[19] J.-B. Huang, A. Singh, and N. Ahuja. Single image super-resolution from transformed self-exemplars. In
CVPR . 1, 3 1720] K. Jia, X. Wang, and X. Tang. Image transformation based on learning dictionaries across image spaces.
IEEE transactions on pattern analysis and machine intelligence , 35(2):367–380, 2013. 1[21] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe:Convolutional architecture for fast feature embedding. In
ACM MM , 2014. 8[22] J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In
ECCV , 2016. 1[23] J. Kim, J. Kwon Lee, and K. Mu Lee. Accurate image super-resolution using very deep convolutionalnetworks. In
CVPR , 2016. 1, 8, 9, 10, 15[24] K. I. Kim and Y. Kwon. Single-image super-resolution using sparse regression and natural image prior.
IEEEtransactions on pattern analysis and machine intelligence , 32(6):1127–1133, 2010. 1[25] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In
ICLR , 2014. 8[26] T. K¨ohler, M. B¨atz, F. Naderi, A. Kaup, A. Maier, and C. Riess. Bridging the simulated-to-real gap: Bench-marking super-resolution on real data. arXiv preprint arXiv:1809.06420 , 2018. 1, 2, 3[27] S. Kong and C. Fowlkes. Image reconstruction with predictive filter flow. arXiv preprint arXiv:1811.11482 ,2018. 7[28] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. nature , 521(7553):436, 2015. 1[29] C. Ledig, L. Theis, F. Husz´ar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang,et al. Photo-realistic single image super-resolution using a generative adversarial network. In
CVPR , 2017.1, 8, 9, 10, 15[30] B. Lim, S. Son, H. Kim, S. Nah, and K. Mu Lee. Enhanced deep residual networks for single image super-resolution. In
CVPRW , pages 136–144, 2017. 1, 7, 8, 13[31] D. G. Lowe. Distinctive image features from scale-invariant keypoints.
International journal of computervision , 60(2):91–110, 2004. 6[32] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its appli-cation to evaluating segmentation algorithms and measuring ecological statistics. In null , page 416. IEEE,2001. 3[33] Y. Matsui, K. Ito, Y. Aramaki, A. Fujimoto, T. Ogawa, T. Yamasaki, and K. Aizawa. Sketch-based mangaretrieval using manga109 dataset.
Multimedia Tools and Applications , 76(20):21811–21838, 2017. 3[34] B. Mildenhall, J. T. Barron, J. Chen, D. Sharlet, R. Ng, and R. Carroll. Burst denoising with kernel predictionnetworks. In
CVPR , 2018. 3, 4, 7, 10[35] S. Niklaus, L. Mai, and F. Liu. Video frame interpolation via adaptive convolution. In
CVPR , 2017. 3, 4[36] S. Niklaus, L. Mai, and F. Liu. Video frame interpolation via adaptive separable convolution. In
ICCV , 2017.3, 4[37] J.-M. Odobez and P. Bouthemy. Robust multiresolution estimation of parametric motion models.
Journal ofvisual communication and image representation , 6(4):348–365, 1995. 6[38] S. C. Park, M. K. Park, and M. G. Kang. Super-resolution image reconstruction: a technical overview.
IEEEsignal processing magazine , 20(3):21–36, 2003. 1[39] C. Qu, D. Luo, E. Monari, T. Schuchert, and J. Beyerer. Capturing ground truth super-resolution data.
ICIP ,2016. 2, 3[40] M. S. Sajjadi, B. Scholkopf, and M. Hirsch. Enhancenet: Single image super-resolution through automatedtexture synthesis. In
ICCV , 2017. 1 1841] S. Schulter, C. Leistner, and H. Bischof. Fast and accurate image upscaling with super-resolution forests. In
CVPR , 2015. 1[42] W. Shi, J. Caballero, F. Husz´ar, J. Totz, A. P. Aitken, R. Bishop, D. Rueckert, and Z. Wang. Real-time singleimage and video super-resolution using an efficient sub-pixel convolutional neural network. In
CVPR . 1, 8,13[43] J. Sun, W. Cao, Z. Xu, and J. Ponce. Learning a convolutional neural network for non-uniform motion blurremoval. In
CVPR , 2015. 3, 4, 7[44] Y. Tai, J. Yang, and X. Liu. Image super-resolution via deep recursive residual network. In
CVPR , 2017. 1[45] R. Timofte, E. Agustsson, L. Van Gool, M.-H. Yang, and L. Zhang. Ntire 2017 challenge on single imagesuper-resolution: Methods and results. In
CVPRW , 2017. 1, 2, 3, 8[46] R. Timofte, V. De Smet, and L. Van Gool. Anchored neighborhood regression for fast example-based super-resolution. In
ICCV , 2013. 1[47] R. Timofte, S. Gu, J. Wu, and L. Van Gool. Ntire 2018 challenge on single image super-resolution: methodsand results. In
CVPRW , 2018. 1, 7, 8[48] T. Vogels, F. Rousselle, B. McWilliams, G. R¨othlin, A. Harvill, D. Adler, M. Meyer, and J. Nov´ak. Denoisingwith kernel prediction and asymmetric loss functions.
ACM Transactions on Graphics (TOG) , 37(4):124,2018. 3, 4, 7, 10[49] Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, et al. Image quality assessment: from error visibilityto structural similarity.
IEEE transactions on image processing , 13(4):600–612, 2004. 8[50] Z. Wang, D. Liu, J. Yang, W. Han, and T. Huang. Deep networks for image super-resolution with sparseprior. In
ICCV , 2015. 1[51] Wikipedia contributors. Circle of confusion — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Circle_of_confusion&oldid=885361501 , 2019.[Online; accessed 6-March-2019]. 7[52] Wikipedia contributors. Depth of field — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Depth_of_field&oldid=886691980 , 2019. [Online; accessed 6-March-2019]. 5, 7[53] Wikipedia contributors. Lens (optics) — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Lens_(optics)&oldid=882234499 , 2019. [Online; accessed 6-March-2019]. 4[54] C.-Y. Yang, C. Ma, and M.-H. Yang. Single-image super-resolution: A benchmark. In
ECCV , 2014. 1[55] J. Yang, Z. Lin, and S. Cohen. Fast image super-resolution based on in-place example regression. In
CVPR .1[56] J. Yang, J. Wright, T. S. Huang, and Y. Ma. Image super-resolution via sparse representation.
IEEE transac-tions on image processing , 19(11):2861–2873, 2010. 1[57] W. Yang, X. Zhang, Y. Tian, W. Wang, and J.-H. Xue. Deep learning for single image super-resolution: Abrief review. arXiv preprint arXiv:1808.03344 , 2018. 1[58] H. Yong, D. Meng, W. Zuo, and L. Zhang. Robust online matrix factorization for dynamic backgroundsubtraction.
IEEE transactions on pattern analysis and machine intelligence , 40(7):1726–1740, 2018. 6[59] R. Zeyde, M. Elad, and M. Protter. On single image scale-up using sparse-representations. In
Internationalconference on curves and surfaces , pages 711–730. Springer, 2010. 3[60] K. Zhang, W. Zuo, and L. Zhang. Learning a single convolutional super-resolution network for multipledegradations. In
CVPR . 2, 3, 8, 13 1961] Y. Zhang, K. Li, K. Li, L. Wang, B. Zhong, and Y. Fu. Image super-resolution using very deep residualchannel attention networks. In
ECCV , 2018. 1, 2, 3, 8, 9, 10, 13, 15[62] Y. Zhang, Y. Tian, Y. Kong, B. Zhong, and Y. Fu. Residual dense network for image super-resolution. In