专栏首页SnailTyanESRGAN - Enhanced Super-Resolution Generative Adversarial Networks论文翻译——中英文对照

ESRGAN - Enhanced Super-Resolution Generative Adversarial Networks论文翻译——中英文对照




ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks


The Super-Resolution Generative Adversarial Network (SR-GAN) [1] is a seminal work that is capable of generating realistic textures during single image super-resolution. However, the hallucinated details are often accompanied with unpleasant artifacts. To further enhance the visual quality, we thoroughly study three key components of SRGAN – network architecture, adversarial loss and perceptual loss, and improve each of them to derive an Enhanced SRGAN (ESRGAN). In particular, we introduce the Residual-in-Residual Dense Block (RRDB) without batch normalization as the basic network building unit. Moreover, we borrow the idea from relativistic GAN [2] to let the discriminator predict relative realness instead of the absolute value. Finally, we improve the perceptual loss by using the features before activation, which could provide stronger supervision for brightness consistency and texture recovery. Benefiting from these improvements, the proposed ESRGAN achieves consistently better visual quality with more realistic and natural textures than SRGAN and won the first place in the PIRM2018-SR Challenge [3]. The code is available at https://github.com/xinntao/ESRGAN.


超分辨率生成对抗网络(SR-GAN)[1]是一项开创性的工作,其能够在单图像超分辨率期间生成逼真的纹理。然而,虚幻的细节常常伴随讨厌的伪像。为了进一步增强视觉质量,我们充分研究了SRGAN的三个关键组成部分——网络架构、对抗损失和感知损失,并对每一个都进行了改进以取得增强的SRGAN(ESRGAN)。特别的是,我们引入了没有批归一化的Residual-in-Residual Dense Block(RRDB)作为基本的网络构架单元。此外,我们借鉴了相对GAN[2]中的思想,让判别器预测相对真实性而不是绝对值。最后,我们通过使用激活前的特征改进感知损失,这可以对亮度一致性和纹理复原提供更强的监督。得益于这些改进,相比于SRGAN,提出的ESRGAN一致地取得了更好的视觉质量、更多真实自然的纹理,并在PIRM2018-SR Challenge[3]中获得了第一名。源码地址:https://github.com/xinntao/ESRGAN

1 Introduction

Single image super-resolution (SISR), as a fundamental low-level vision problem, has attracted increasing attention in the research community and AI companies. SISR aims at recovering a high-resolution (HR) image from a single low-resolution (LR) one. Since the pioneer work of SRCNN proposed by Dong et al. [4], deep convolution neural network (CNN) approaches have brought prosperous development. Various network architecture designs and training strategies have continuously improved the SR performance, especially the Peak Signal-toNoise Ratio (PSNR) value [5,6,7,1,8,9,10,11,12]. However, these PSNR-oriented approaches tend to output over-smoothed results without sufficient high-frequency details, since the PSNR metric fundamentally disagrees with the subjective evaluation of human observers [1].

1 引言


Several perceptual-driven methods have been proposed to improve the visual quality of SR results. For instance, perceptual loss [13,14] is proposed to optimize super-resolution model in a feature space instead of pixel space. Generative adversarial network [15] is introduced to SR by [1,16] to encourage the network to favor solutions that look more like natural images. The semantic image prior is further incorporated to improve recovered texture details [17]. One of the milestones in the way pursuing visually pleasing results is SRGAN [1]. The basic model is built with residual blocks [18] and optimized using perceptual loss in a GAN framework. With all these techniques, SRGAN significantly improves the overall visual quality of reconstruction over PSNR-oriented methods.


However, there still exists a clear gap between SRGAN results and the ground-truth (GT) images, as shown in Fig. 1. In this study, we revisit the key components of SRGAN and improve the model in three aspects. First, we improve the network structure by introducing the Residual-in-Residual Dense Block (RDDB), which is of higher capacity and easier to train. We also remove Batch Normalization (BN) [19] layers as in [20] and use residual scaling [21,20] and smaller initialization to facilitate training a very deep network. Second, we improve the discriminator using Relativistic average GAN (RaGAN) [2], which learns to judge “whether one image is more realistic than the other” rather than “whether one image is real or fake”. Our experiments show that this improvement helps the generator recover more realistic texture details. Third, we propose an improved perceptual loss by using the VGG features before activation instead of after activation as in SRGAN. We empirically find that the adjusted perceptual loss provides sharper edges and more visually pleasing results, as will be shown in Sec. 4.4. Extensive experiments show that the enhanced SRGAN, termed ESRGAN, consistently outperforms state-of-the-art methods in both sharpness and details (see Fig. 1 and Fig. 7).

Fig.1: The super-resolution results of ×4 for SRGAN, the proposed ESRGAN and the ground-truth. ESRGAN outperforms SRGAN in sharpness and details.

Fig.7: Qualitative results of ESRGAN. ESRGAN produces more natural textures, e.g., animal fur, building structure and grass texture, and also less unpleasant artifacts, e.g., artifacts in the face by SRGAN.

然而,如图1所示,SRGAN结果与真实(GT)图像之间仍然存在明显的差距。在本研究中,我们重新审视SRGAN的关键组件,并在三个方面改进模型。首先,我们通过引入Residual-in-Residual Dense Block(RDDB)改进网络架构,该结构具有较高的能力且更容易训练。我们像[20]中一样也移除了批归一化(BN)[19]层,使用残差缩放[21,20]和更小的初始化来促进训练一个非常深的网络。其次,我们使用相对平均GAN(RaGAN)[2]来改进判别器,RaGAN学习判断“一张图像是否比另一张更真实”而不是“一张图像时真的还是假的”。我们的实验表明这个改进有助于生成器恢复更多的真实纹理细节。第三,我们提出了一种改进的感知损失,使用激活之前的VGG特征来代替SRGAN中激活之后的VGG特征。从经验上我们发现调整之后的感知损失提供了更清晰的边缘和视觉上更令人满意的结果,如4.4节所示。大量的实验表明增强SRGAN(称为ESRGAN)在清晰度和细节方面都始终优于最新的方法(见图1和图7)。



We take a variant of ESRGAN to participate in the PIRM-SR Challenge [3]. This challenge is the first SR competition that evaluates the performance in a perceptual-quality aware manner based on [22], where the authors claim that distortion and perceptual quality are at odds with each other. The perceptual quality is judged by the non-reference measures of Ma’s score [23] and NIQE [24], i.e., perceptual index =12((10−Ma)+NIQE)=\frac {1} {2}((10−Ma)+NIQE)=21​((10−Ma)+NIQE). A lower perceptual index represents a better perceptual quality.

我们采用ESRGAN的一个变种来参加PIRM-SR挑战赛[3]。这个挑战是第一个在[22]的基础上以察觉感知质量的方式评估性能的SR竞赛,[22]中作者声称失真和感知质量相互矛盾。感知质量是通过Ma分数[23]和NIQE[24]的非参考度量来判断的,即感知指数=12((10−Ma)+NIQE)=\frac {1} {2}((10−Ma)+NIQE)=21​((10−Ma)+NIQE)。更低的感知指数表示更好的感知质量。

As shown in Fig. 2, the perception-distortion plane is divided into three regions defined by thresholds on the Root-Mean-Square Error (RMSE), and the algorithm that achieves the lowest perceptual index in each region becomes the regional champion. We mainly focus on region 3 as we aim to bring the perceptual quality to a new high. Thanks to the aforementioned improvements and some other adjustments as discussed in Sec. 4.6, our proposed ESRGAN won the first place in the PIRM-SR Challenge (region 3) with the best perceptual index.

Fig.2: Perception-distortion plane on PIRM self validation dataset. We show the baselines of EDSR [20], RCAN [12] and EnhanceNet [16], and the submitted ESRGAN model. The blue dots are produced by image interpolation.



In order to balance the visual quality and RMSE/PSNR, we further propose the network interpolation strategy, which could continuously adjust the reconstruction style and smoothness. Another alternative is image interpolation, which directly interpolates images pixel by pixel. We employ this strategy to participate in region 1 and region 2. The network interpolation and image interpolation strategies and their differences are discussed in Sec. 3.4.


We focus on deep neural network approaches to solve the SR problem. As a pioneer work, Dong et al. [4,25] propose SRCNN to learn the mapping from LR to HR images in an end-to-end manner, achieving superior performance against previous works. Later on, the field has witnessed a variety of network architectures, such as a deeper network with residual learning [5], Laplacian pyramid structure [6], residual blocks [1], recursive learning [7,8], densely connected network [9], deep back projection [10] and residual dense network [11]. Specifically, Lim et al. [20] propose EDSR model by removing unnecessary BN layers in the residual block and expanding the model size, which achieves significant improvement. Zhang et al. [11] propose to use effective residual dense block in SR, and they further explore a deeper network with channel attention [12], achieving the state-of-the-art PSNR performance. Besides supervised learning, other methods like reinforcement learning [26] and unsupervised learning [27] are also introduced to solve general image restoration problems.

2 相关工作


Several methods have been proposed to stabilize training a very deep model. For instance, residual path is developed to stabilize the training and improve the performance [18,5,12]. Residual scaling is first employed by Szegedy et al. [21] and also used in EDSR. For general deep networks, He et al. [28] propose a robust initialization method for VGG-style networks without BN. To facilitate training a deeper network, we develop a compact and effective residual-in-residual dense block, which also helps to improve the perceptual quality.


Perceptual-driven approaches have also been proposed to improve the visual quality of SR results. Based on the idea of being closer to perceptual similarity [29,14] perceptual loss [13] is proposed to enhance the visual quality by minimizing the error in a feature space instead of pixel space. Contextual loss [30] is developed to generate images with natural image statistics by using an objective that focuses on the feature distribution rather than merely comparing the appearance. Ledig et al. [1] propose SRGAN model that uses perceptual loss and adversarial loss to favor outputs residing on the manifold of natural images. Sajjadi et al. [16] develop a similar approach and further explored the local texture matching loss. Based on these works, Wang et al. [17] propose spatial feature transform to effectively incorporate semantic prior in an image and improve the recovered textures.


Throughout the literature, photo-realism is usually attained by adversarial training with GAN [15]. Recently there are a bunch of works that focus on developing more effective GAN frameworks. WGAN [31] proposes to minimize a reasonable and efficient approximation of Wasserstein distance and regularizes discriminator by weight clipping. Other improved regularization for discriminator includes gradient clipping [32] and spectral normalization [33]. Relativistic discriminator [2] is developed not only to increase the probability that generated data are real, but also to simultaneously decrease the probability that real data are real. In this work, we enhance SRGAN by employing a more effective relativistic average GAN.


SR algorithms are typically evaluated by several widely used distortion measures, e.g., PSNR and SSIM. However, these metrics fundamentally disagree with the subjective evaluation of human observers [1]. Non-reference measures are used for perceptual quality evaluation, including Ma’s score [23] and NIQE [24], both of which are used to calculate the perceptual index in the PIRM-SR Challenge [3]. In a recent study, Blau et al. [22] find that the distortion and perceptual quality are at odds with each other.


3 Proposed Methods

Our main aim is to improve the overall perceptual quality for SR. In this section, we first describe our proposed network architecture and then discuss the improvements from the discriminator and perceptual loss. At last, we describe the network interpolation strategy for balancing perceptual quality and PSNR.

3 提出的方法


3.1 Network Architecture

In order to further improve the recovered image quality of SRGAN, we mainly make two modifications to the structure of generator G: 1) remove all BN layers; 2) replace the original basic block with the proposed Residual-in-Residual Dense Block (RRDB), which combines multi-level residual network and dense connections as depicted in Fig. 4.

Fig.4: Left: We remove the BN layers in residual block in SRGAN. Right: RRDB block is used in our deeper model and β\betaβ is the residual scaling parameter.

3.1 网络架构



Removing BN layers has proven to increase performance and reduce computational complexity in different PSNR-oriented tasks including SR [20] and deblurring [35]. BN layers normalize the features using mean and variance in a batch during training and use estimated mean and variance of the whole training dataset during testing. When the statistics of training and testing datasets differ a lot, BN layers tend to introduce unpleasant artifacts and limit the generalization ability. We empirically observe that BN layers are more likely to bring artifacts when the network is deeper and trained under a GAN framework. These artifacts occasionally appear among iterations and different settings, violating the needs for a stable performance over training. We therefore remove BN layers for stable training and consistent performance. Furthermore, removing BN layers helps to improve generalization ability and to reduce computational complexity and memory usage.


We keep the high-level architecture design of SRGAN (see Fig. 3), and use a novel basic block namely RRDB as depicted in Fig. 4. Based on the observation that more layers and connections could always boost performance [20,11,12], the proposed RRDB employs a deeper and more complex structure than the original residual block in SRGAN. Specifically, as shown in Fig. 4, the proposed RRDB has a residual-in-residual structure, where residual learning is used in different levels. A similar network structure is proposed in [36] that also applies a multilevel residual network. However, our RRDB differs from [36] in that we use dense block [34] in the main path as [11], where the network capacity becomes higher benefiting from the dense connections.

Fig. 3: We employ the basic architecture of SRResNet [1], where most computation is done in the LR feature space. We could select or design “basic blocks” (e.g., residual block [18], dense block [34], RRDB) for better performance.



In addition to the improved architecture, we also exploit several techniques to facilitate training a very deep network: 1) residual scaling [21,20], i.e., scaling down the residuals by multiplying a constant between 0 and 1 before adding them to the main path to prevent instability; 2) smaller initialization, as we empirically find residual architecture is easier to train when the initial parameter variance becomes smaller. More discussion can be found in the supplementary material.


The training details and the effectiveness of the proposed network will be presented in Sec. 4.


3.2 Relativistic Discriminator

Besides the improved structure of generator, we also enhance the discriminator based on the Relativistic GAN [2]. Different from the standard discriminator DDD in SRGAN, which estimates the probability that one input image xxx is real and natural, a relativistic discriminator tries to predict the probability that a real image xrx_rxr​ is relatively more realistic than a fake one xfx_fxf​ , as shown in Fig. 5.

Fig. 5: Difference between standard discriminator and relativistic discriminator.

3.2 相对判别器



Specifically, we replace the standard discriminator with the Relativistic average Discriminator RaD [2], denoted as DRaD_{Ra}DRa​. The standard discriminator in SRGAN can be expressed as D(x)=σ(C(x))D(x) = \sigma(C(x))D(x)=σ(C(x)), where σ\sigmaσ is the sigmoid function and C(x)C(x)C(x) is the non-transformed discriminator output. Then the RaD is formulated as KaTeX parse error: Undefined control sequence: \[ at position 52: …mathbb{E}_{x_f}\̲[̲C(x_f)\]), where KaTeX parse error: Undefined control sequence: \[ at position 17: …mathbb{E}_{x_f}\̲[̲\bullet\] represents the operation of taking average for all fake data in the mini-batch. The discriminator loss is then defined as: KaTeX parse error: Undefined control sequence: \[ at position 30: …mathbb{E}_{x_r}\̲[̲log(D_{Ra}(x_r,…

The adversarial loss for generator is in a symmetrical form: KaTeX parse error: Undefined control sequence: \[ at position 30: …mathbb{E}_{x_r}\̲[̲1-log(D_{Ra}(x_…

where xf=G(xi)x_f = G(x_i)xf​=G(xi​) and xix_ixi​ stands for the input LR image. It is observed that the adversarial loss for generator contains both xrx_rxr​ and xfx_fxf​. Therefore, our generator benefits from the gradients from both generated data and real data in adversarial training, while in SRGAN only generated part takes effect. In Sec. 4.4, we will show that this modification of discriminator helps to learn sharper edges and more detailed textures.

具体来说,我们用相对平均判别器RaD[2]代替标准判别器,记为DRaD_{Ra}DRa​。SRGAN中的标准判别器可表示为D(x)=σ(C(x))D(x) = \sigma(C(x))D(x)=σ(C(x)),其中σ\sigmaσ是sigmoid函数,C(x)C(x)C(x)是非变换判别器输出。然后RaD用公式表示为KaTeX parse error: Undefined control sequence: \[ at position 52: …mathbb{E}_{x_f}\̲[̲C(x_f)\]),其中KaTeX parse error: Undefined control sequence: \[ at position 17: …mathbb{E}_{x_f}\̲[̲\bullet\]表示对小批次中所有假数据取平均值的操作。然后判别器损失定义为:KaTeX parse error: Undefined control sequence: \[ at position 30: …mathbb{E}_{x_r}\̲[̲log(D_{Ra}(x_r,…

生成器的对抗损失呈对称形式:KaTeX parse error: Undefined control sequence: \[ at position 30: …mathbb{E}_{x_r}\̲[̲1-log(D_{Ra}(x_…


3.3 Perceptual Loss

We also develop a more effective perceptual loss LpercepL_{percep}Lpercep​ by constraining on features before activation rather than after activation as practiced in SRGAN.

3.3 感知损失


Based on the idea of being closer to perceptual similarity [29,14], Johnson et al. [13] propose perceptual loss and it is extended in SRGAN [1]. Perceptual loss is previously defined on the activation layers of a pre-trained deep network, where the distance between two activated features is minimized. Contrary to the convention, we propose to use features before the activation layers, which will overcome two drawbacks of the original design. First, the activated features are very sparse, especially after a very deep network, as depicted in Fig. 6. For example, the average percentage of activated neurons for image ‘baboon’ after VGG19-54 layer is merely 11.17%11.17\%11.17%. The sparse activation provides weak supervision and thus leads to inferior performance. Second, using features after activation also causes inconsistent reconstructed brightness compared with the ground-truth image, which we will show in Sec. 4.4.

Fig.6: Representative feature maps before and after activation for image ‘baboon’. With the network going deeper, most of the features after activation become inactive while features before activation contains more information.



Therefore, the total loss for the generator is: LG=Lpercep+λLGRa+ηL1(3)L_G = L_{percep} + \lambda L^{Ra}_G + \eta L_1 \tag{3}LG​=Lpercep​+λLGRa​+ηL1​(3) where L1=Exi∣∣G(xi)−y∣∣1L_1 = \mathbb{E}_{x_i} || G(x_i) − y||_1L1​=Exi​​∣∣G(xi​)−y∣∣1​ is the content loss that evaluate the 1-norm distance between recovered image G(xi)G(x_i)G(xi​) and the ground-truth yyy, and λ,η\lambda, \etaλ,η are the coefficients to balance different loss terms.

因此,生成器的全部损失为:LG=Lpercep+λLGRa+ηL1(3)L_G = L_{percep} + \lambda L^{Ra}_G + \eta L_1 \tag{3}LG​=Lpercep​+λLGRa​+ηL1​(3),其中L1=Exi∣∣G(xi)−y∣∣1L_1 = \mathbb{E}_{x_i} || G(x_i) − y||_1L1​=Exi​​∣∣G(xi​)−y∣∣1​是内容损失,用来评估恢复图像G(xi)G(x_i)G(xi​)和真实图像yyy之间的1范数距离,λ,η\lambda, \etaλ,η是平衡不同损失项的系数。

We also explore a variant of perceptual loss in the PIRM-SR Challenge. In contrast to the commonly used perceptual loss that adopts a VGG network trained for image classification, we develop a more suitable perceptual loss for SR–MINC loss. It is based on a fine-tuned VGG network for material recognition [38], which focuses on textures rather than object. Although the gain of perceptual index brought by MINC loss is marginal, we still believe that exploring perceptual loss that focuses on texture is critical for SR.


3.4 Network Interpolation

To remove unpleasant noise in GAN-based methods while maintain a good perceptual quality, we propose a flexible and effective strategy – network interpolation. Specifically, we first train a PSNR-oriented network GPSNRG_{PSNR}GPSNR​ and then obtain a GAN-based network GGANG_{GAN}GGAN​ by fine-tuning. We interpolate all the corresponding parameters of these two networks to derive an interpolated model GINTERPG_{INTERP}GINTERP​, whose parameters are: θGINTERP=(1−α)θGPSNR+αθGGAN(4)\theta^{INTERP}_{G} = (1 − \alpha) \theta^{PSNR}_{G} + \alpha \theta^{GAN}_{G} \tag{4}θGINTERP​=(1−α)θGPSNR​+αθGGAN​(4) where GINTERPG_{INTERP}GINTERP​, GPSNRG_{PSNR}GPSNR​ and GGANG_{GAN}GGAN​ are the parameters of θGINTERP\theta^{INTERP}_{G}θGINTERP​, θGPSNR\theta^{PSNR}_{G}θGPSNR​ and θGGAN\theta^{GAN}_{G}θGGAN​, respectively, and α∈[0,1]\alpha \in [0, 1]α∈[0,1] is the interpolation parameter.

3.4 网络插值

为了去除基于GAN方法中讨厌的噪声同时保持好的感知质量,我们提出了一种弹性有效的策略——网络插值。具体来说,我们首先训练一个面向PSNR的网络GPSNRG_{PSNR}GPSNR​,然后通过微调获得一个基于GAN的网络GGANG_{GAN}GGAN​。我们插值这两个网络的所有对应参数来取得插值模型GINTERPG_{INTERP}GINTERP​,其参数为:θGINTERP=(1−α)θGPSNR+αθGGAN(4)\theta^{INTERP}_{G} = (1 − \alpha) \theta^{PSNR}_{G} + \alpha \theta^{GAN}_{G} \tag{4}θGINTERP​=(1−α)θGPSNR​+αθGGAN​(4) 其中GINTERPG_{INTERP}GINTERP​, GPSNRG_{PSNR}GPSNR​和GGANG_{GAN}GGAN​分别是θGINTERP\theta^{INTERP}_{G}θGINTERP​, θGPSNR\theta^{PSNR}_{G}θGPSNR​和θGGAN\theta^{GAN}_{G}θGGAN​的参数,α∈[0,1]\alpha \in [0, 1]α∈[0,1]为插值参数。

The proposed network interpolation enjoys two merits. First, the interpolated model is able to produce meaningful results for any feasible α\alphaα without introducing artifacts. Second, we can continuously balance perceptual quality and fidelity without re-training the model.


We also explore alternative methods to balance the effects of PSNR-oriented and GAN-based methods. For instance, one can directly interpolate their output images (pixel by pixel) rather than the network parameters. However, such an approach fails to achieve a good trade-off between noise and blur, i.e., the interpolated image is either too blurry or noisy with artifacts (see Sec. 4.5). Another method is to tune the weights of content loss and adversarial loss, i.e., the parameter λ\lambdaλ and η\etaη in Eq. (3). But this approach requires tuning loss weights and fine-tuning the network, and thus it is too costly to achieve continuous control of the image style.


4 Experiments

4.1 Training Details

Following SRGAN [1], all experiments are performed with a scaling factor of ×4 between LR and HR images. We obtain LR images by down-sampling HR images using the MATLAB bicubic kernel function. The mini-batch size is set to 16. The spatial size of cropped HR patch is 128 × 128. We observe that training a deeper network benefits from a larger patch size, since an enlarged receptive field helps to capture more semantic information. However, it costs more training time and consumes more computing resources. This phenomenon is also observed in PSNR-oriented methods (see supplementary material).

4 实验

4.1 训练细节


The training process is divided into two stages. First, we train a PSNR-oriented model with the L1 loss. The learning rate is initialized as 2×10−42 × 10^{−4}2×10−4 and decayed by a factor of 2 every 2×1052 × 10^52×105 of mini-batch updates. We then employ the trained PSNR-oriented model as an initialization for the generator. The generator is trained using the loss function in Eq. (3) with λ=5×10−3\lambda = 5×10^{−3}λ=5×10−3 and η=1×10−2\eta = 1×10^{−2}η=1×10−2. The learning rate is set to 1×10−41×10^{−4}1×10−4 and halved at [50k, 100k, 200k, 300k] iterations. Pre-training with pixel-wise loss helps GAN-based methods to obtain more visually pleasing results. The reasons are that 1) it can avoid undesired local optima for the generator; 2) after pre-training, the discriminator receives relatively good super-resolved images instead of extreme fake ones (black or noisy images) at the very beginning, which helps it to focus more on texture discrimination.

训练过程分为两个阶段。首先,我们训练一个具有L1损失的面向PSNR的模型。学习率初始化为2×10−42 × 10^{−4}2×10−4,每2×1052 × 10^52×105个小批次更新的衰减因子为2。然后,我们采用训练的面向PSNR的模型作为生成器的初始化。生成器训练使用等式3中的损失函数,λ=5×10−3\lambda = 5×10^{−3}λ=5×10−3,η=1×10−2\eta = 1×10^{−2}η=1×10−2。学习率设置为1×10−41×10^{−4}1×10−4,在[50k, 100k, 200k, 300k]次迭代之后减半。使用逐像素损失进行预训练有助于基于GAN的方法获得视觉上更好的结果。原因是:1)它可以避免生成器不希望的局部最优;2)在预训练之后,最初判别器可以收到相对好的超分辨率图像而不是极端假的图像(黑色或噪声图像),这有助于其更关注纹理判别。

For optimization, we use Adam [39] with β1=0.9,β2=0.999\beta_1 = 0.9, \beta_2 = 0.999β1​=0.9,β2​=0.999. We alternately update the generator and discriminator network until the model converges. We use two settings for our generator – one of them contains 16 residual blocks, with a capacity similar to that of SRGAN and the other is a deeper model with 23 RRDB blocks. We implement our models with the PyTorch framework and train them using NVIDIA Titan Xp GPUs.

为了优化,我们使用Adam[39],其中β1=0.9,β2=0.999\beta_1 = 0.9, \beta_2 = 0.999β1​=0.9,β2​=0.999。我们交替更新生成器和判别器网络,直到模型收敛。我们为生成器使用了两种设置——其中一种包含16个残差块,能力类似于SRGAN,另一种是具有23个RRDB块的更深的模型。我们使用PyTorch框架实现我们的模型,并使用NVIDIA Titan Xp GPU对其进行训练。

4.2 Data

For training, we mainly use the DIV2K dataset [40], which is a high-quality (2K resolution) dataset for image restoration tasks. Beyond the training set of DIV2K that contains 800 images, we also seek for other datasets with rich and diverse textures for our training. To this end, we further use the Flickr2K dataset [41] consisting of 2650 2K high-resolution images collected on the Flickr website, and the OutdoorSceneTraining (OST) [17] dataset to enrich our training set. We empirically find that using this large dataset with richer textures helps the generator to produce more natural results, as shown in Fig. 8.

Fig. 8: Overall visual comparisons for showing the effects of each component in ESRGAN. Each column represents a model with its configurations in the top. The red sign indicates the main improvement compared with the previous model.

4.2 数据



We train our models in RGB channels and augment the training dataset with random horizontal flips and 90 degree rotations. We evaluate our models on widely used benchmark datasets – Set5 [42], Set14 [43], BSD100 [44], Urban100 [45], and the PIRM self-validation dataset that is provided in the PIRM-SR Challenge.


4.3 Qualitative Results

We compare our final models on several public benchmark datasets with state-ofthe-art PSNR-oriented methods including SRCNN [4], EDSR [20] and RCAN [12], and also with perceptual-driven approaches including SRGAN [1] and EnhanceNet [16]. Since there is no effective and standard metric for perceptual quality, we present some representative qualitative results in Fig. 7. PSNR (evaluated on the luminance channel in YCbCr color space) and the perceptual index used in the PIRM-SR Challenge are also provided for reference.

4.3 定性结果


It can be observed from Fig. 7 that our proposed ESRGAN outperforms previous approaches in both sharpness and details. For instance, ESRGAN can produce sharper and more natural baboon’s whiskers and grass textures (see image 43074) than PSNR-oriented methods, which tend to generate blurry results, and than previous GAN-based methods, whose textures are unnatural and contain unpleasing noise. ESRGAN is capable of generating more detailed structures in building (see image 102061) while other methods either fail to produce enough details (SRGAN) or add undesired textures (EnhanceNet). Moreover, previous GAN-based methods sometimes introduce unpleasant artifacts, e.g., SRGAN adds wrinkles to the face. Our ESRGAN gets rid of these artifacts and produces natural results.


4.4 Ablation Study

In order to study the effects of each component in the proposed ESRGAN, we gradually modify the baseline SRGAN model and compare their differences. The overall visual comparison is illustrated in Fig. 8. Each column represents a model with its configurations shown in the top. The red sign indicates the main improvement compared with the previous model. A detailed discussion is provided as follows.

4.4 消融研究


BN removal. We first remove all BN layers for stable and consistent performance without artifacts. It does not decrease the performance but saves the computational resources and memory usage. For some cases, a slight improvement can be observed from the 2nd and 3rd columns in Fig. 8 (e.g., image 39). Furthermore, we observe that when a network is deeper and more complicated, the model with BN layers is more likely to introduce unpleasant artifacts. The examples can be found in the supplementary material.


Before activation in perceptual loss. We first demonstrate that using features before activation can result in more accurate brightness of reconstructed images. To eliminate the influences of textures and color, we filter the image with a Gaussian kernel and plot the histogram of its gray-scale counterpart. Fig. 9a shows the distribution of each brightness value. Using activated features skews the distribution to the left, resulting in a dimmer output while using features before activation leads to a more accurate brightness distribution closer to that of the ground-truth.

Fig. 9: Comparison between before activation and after activation.



We can further observe that using features before activation helps to produce sharper edges and richer textures as shown in Fig. 9b (see bird feather) and Fig. 8 (see the 3rd and 4th columns), since the dense features before activation offer a stronger supervision than that a sparse activation could provide.


RaGAN. RaGAN uses an improved relativistic discriminator, which is shown to benefit learning sharper edges and more detailed textures. For example, in the 5th column of Fig. 8, the generated images are sharper with richer textures than those on their left (see the baboon, image 39 and image 43074).


Deeper network with RRDB. Deeper model with the proposed RRDB can further improve the recovered textures, especially for the regular structures like the roof of image 6 in Fig. 8, since the deep model has a strong representation capacity to capture semantic information. Also, we find that a deeper model can reduce unpleasing noises like image 20 in Fig. 8.

具有RRDB的更深网络。具有提出的RRDB的更深模型可以进一步改善恢复的纹理,尤其是像图8中图像6的屋顶这样的常规结构,因为深度模型具有强大的表示能力来捕获语义信息。 我们也发现更深的模型可以减少像图8中图像20这样的令人不快的噪声。

In contrast to SRGAN, which claimed that deeper models are increasingly difficult to train, our deeper model shows its superior performance with easy training, thanks to the improvements mentioned above especially the proposed RRDB without BN layers.


4.5 Network Interpolation

We compare the effects of network interpolation and image interpolation strategies in balancing the results of a PSNR-oriented model and GAN-based method. We apply simple linear interpolation on both the schemes. The interpolation parameter α\alphaα is chosen from 0 to 1 with an interval of 0.2.

4.5 网络插值


As depicted in Fig. 10, the pure GAN-based method produces sharp edges and richer textures but with some unpleasant artifacts, while the pure PSNRoriented method outputs cartoon-style blurry images. By employing network interpolation, unpleasing artifacts are reduced while the textures are maintained. By contrast, image interpolation fails to remove these artifacts effectively.

Fig. 10: The comparison between network interpolation and image interpolation.



Interestingly, it is observed that the network interpolation strategy provides a smooth control of balancing perceptual quality and fidelity in Fig. 10.


4.6 The PIRM-SR Challenge

We take a variant of ESRGAN to participate in the PIRM-SR Challenge [3]. Specifically, we use the proposed ESRGAN with 16 residual blocks and also empirically make some modifications to cater to the perceptual index. 1) The MINC loss is used as a variant of perceptual loss, as discussed in Sec. 3.3. Despite the marginal gain on the perceptual index, we still believe that exploring perceptual loss that focuses on texture is crucial for SR. 2) Pristine dataset [24], which is used for learning the perceptual index, is also employed in our training; 3) a high weight of loss L1L_1L1​ up to η=10\eta = 10η=10 is used due to the PSNR constraints; 4) we also use back projection [46] as post-processing, which can improve PSNR and sometimes lower the perceptual index.

4.6 PIRM-SR挑战赛

我们采用ESRGAN的一个变种来参加PIRM-SR挑战赛[3]。具体来说,我们使用提出的具有16个残差块的ESRGAN,并根据经验进行了一些修改来迎合感知指数。1)使用MINC损失作为感知损失的一个变种,如3.3节所述。尽管在感知指数上有边际收益,但我们仍认为采用专注于纹理的感知损失对于SR至关重要;2)我们的训练中也使用了Pristine数据集[24]来学习感知指数;3)由于PSNR约束,L1L_1L1​损失的权重高达η=10\eta = 10η=10;4)我们也使用反向投射[46]作为后处理,其可以改善PSNR,有时会降低感知指数。

For other regions 1 and 2 that require a higher PSNR, we use image interpolation between the results of our ESRGAN and those of a PSNR-oriented method RCAN [12]. The image interpolation scheme achieves a lower perceptual index (lower is better) although we observed more visually pleasing results by using the network interpolation scheme. Our proposed ESRGAN model won the first place in the PIRM-SR Challenge (region 3) with the best perceptual index.


5 Conclusion

We have presented an ESRGAN model that achieves consistently better perceptual quality than previous SR methods. The method won the first place in the PIRM-SR Challenge in terms of the perceptual index. We have formulated a novel architecture containing several RDDB blocks without BN layers. In addition, useful techniques including residual scaling and smaller initialization are employed to facilitate the training of the proposed deep model. We have also introduced the use of relativistic GAN as the discriminator, which learns to judge whether one image is more realistic than another, guiding the generator to recover more detailed textures. Moreover, we have enhanced the perceptual loss by using the features before activation, which offer stronger supervision and thus restore more accurate brightness and realistic textures.

5 结论


Acknowledgement. This work is supported by SenseTime Group Limited, the General Research Fund sponsored by the Research Grants Council of the Hong Kong SAR (CUHK 14241716, 14224316. 14209217), National Natural Science Foundation of China (U1613211) and Shenzhen Research Program (JCYJ20170818164704758, JCYJ20150925163005055).

致谢。这项工作由商汤科技支持,香港特别行政区研究资助局(CUHK 14241716、14224316、14209217),中国国家自然科学基金(U1613211)和深圳研究计划(JCYJ20170818164704758,JCYJ20150925163005055)赞助。


  1. Ledig,C.,Theis,L.,Husza ́r,F.,Caballero,J.,Cunningham,A.,Acosta,A.,Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image super-resolution using a generative adversarial network. In: CVPR. (2017)
  2. Jolicoeur-Martineau, A.: The relativistic discriminator: a key element missing from standard gan. arXiv preprint arXiv:1807.00734 (2018)
  3. Blau, Y., Mechrez, R., Timofte, R., Michaeli, T., Zelnik-Manor, L.: The pirm challenge on perceptual super resolution. https://www.pirm2018.org/PIRM-SR. html (2018)
  4. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV. (2014)
  5. Kim, J., Kwon Lee, J., Mu Lee, K.: Accurate image super-resolution using very deep convolutional networks. In: CVPR. (2016)
  6. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Deep laplacian pyramid networks for fast and accurate super-resolution. In: CVPR. (2017)
  7. Kim, J., Kwon Lee, J., Mu Lee, K.: Deeply-recursive convolutional network for image super-resolution. In: CVPR. (2016)
  8. Tai, Y., Yang, J., Liu, X.: Image super-resolution via deep recursive residual network. In: CVPR. (2017)
  9. Tai, Y., Yang, J., Liu, X., Xu, C.: Memnet: A persistent memory network for image restoration. In: ICCV. (2017)
  10. Haris, M., Shakhnarovich, G., Ukita, N.: Deep backprojection networks for super- resolution. In: CVPR. (2018)
  11. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR. (2018)
  12. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV. (2018)
  13. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV. (2016)
  14. Bruna, J., Sprechmann, P., LeCun, Y.: Super-resolution with deep convolutional sufficient statistics. In: ICLR. (2015)
  15. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NIPS. (2014)
  16. Sajjadi, M.S., Scho ̈lkopf, B., Hirsch, M.: Enhancenet: Single image super-resolution through automated texture synthesis. In: ICCV. (2017)
  17. Wang, X., Yu, K., Dong, C., Loy, C.C.: Recovering realistic texture in image super-resolution by deep spatial feature transform. In: CVPR. (2018)
  18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. (2016)
  19. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: ICMR. (2015)
  20. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPRW. (2017)
  21. Szegedy, C., Ioffe, S., Vanhoucke, V.: Inception-v4, inception-resnet and the impact of residual connections on learning. arXiv preprint arXiv:1602.07261 (2016)
  22. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: CVPR. (2017)
  23. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. CVIU 158 (2017) 1–16
  24. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a completely blind image quality analyzer. IEEE Signal Process. Lett. 20(3) (2013) 209–212
  25. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. TPAMI 38(2) (2016) 295–307
  26. Yu, K., Dong, C., Lin, L., Loy, C.C.: Crafting a toolchain for image restoration by deep reinforcement learning. In: CVPR. (2018)
  27. Yuan, Y., Liu, S., Zhang, J., Zhang, Y., Dong, C., Lin, L.: Unsupervised image super-resolution using cycle-in-cycle generative adversarial networks. In: CVPRW. (2018)
  28. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: ICCV. (2015)
  29. Gatys, L., Ecker, A.S., Bethge, M.: Texture synthesis using convolutional neural networks. In: NIPS. (2015)
  30. Mechrez, R., Talmi, I., Shama, F., Zelnik-Manor, L.: Maintaining natural image statistics with the contextual loss. arXiv preprint arXiv:1803.04626 (2018)
  31. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein gan. arXiv preprint arXiv:1701.07875 (2017)
  32. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. In: NIPS. (2017)
  33. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)
  34. Huang, G., Liu, Z., Weinberger, K.Q., van der Maaten, L.: Densely connected convolutional networks. In: CVPR. (2017)
  35. Nah, S., Kim, T.H., Lee, K.M.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR. (2017)
  36. Zhang, K., Sun, M., Han, X., Yuan, X., Guo, L., Liu, T.: Residual networks of residual networks: Multilevel residual networks. IEEE Transactions on Circuits and Systems for Video Technology (2017)
  37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)
  38. Bell, S., Upchurch, P., Snavely, N., Bala, K.: Material recognition in the wild with the materials in context database. In: CVPR. (2015)
  39. Kingma, D., Ba, J.: Adam: A method for stochastic optimization. In: ICLR. (2015)
  40. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: CVPRW. (2017)
  41. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L., Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M., et al.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: CVPRW. (2017)
  42. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC, BMVA press (2012)
  43. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: International Conference on Curves and Surfaces, Springer (2010)
  44. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV. (2001)
  45. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR. (2015)
  46. Timofte, R., Rothe, R., Van Gool, L.: Seven ways to improve example-based single image super resolution. In: CVPR. (2016)
  47. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: International Conference on Artificial Intelligence and Statistics. (2010)



0 条评论
登录 后参与评论


  • Inception-V3论文翻译——中英文对照

    Rethinking the Inception Architecture for Computer Vision Abstract Convolutional...

  • Very Deep Convolutional Networks for Large-Scale Image Recognition—VGG论文翻译—中英文对照

    Very Deep Convolutional Networks for Large-Scale Image Recognition ABSTRACT In t...

  • Deformable Convolutional Networks论文翻译——中英文对照

    Deformable Convolutional Networks Abstract Convolutional neural networks (CNNs) ...

  • 使用jMeter测试通过SAP ID Service认证的SAP Cloud API

    Recently I am working on a co-innovation project with one local partner in China...

    Jerry Wang
  • WHY CHINA’S CROSS-BORDER E-COMMERCE MARKET IS SURGING? 2020 Report on Asia-Pacific Cross-Border B...

    At $5.8 trillion last year, China has the second-largest retail market in the wo...

  • SAP JCO的下载,安装,配置和使用

    This blog can be used as a very simple tutorial for those guys who are new to JC...

    Jerry Wang
  • 如何让SAP C4C自定义BO实现附件上传的功能

    I began to learn how to use Cloud Application Studio today and I record what I h...

    Jerry Wang
  • 使用Chrome开发者工具研究JavaScript的垃圾回收机制

    I use the following simple JavaScript code to illustrate:

    Jerry Wang
  • Protocol Buffer的使用(p

    安装完毕后,进入解压目录的python目录,执行python setup.py install;安装python的protobuf库即可。

  • silverlight寻奇 - Graphite

    Graphite是一个能自动布局的图表控件。 目前它已经有了silverlight 2 和 wpf的版本。观看demo时按下“Ctrl”键再做点击操作。 原文...