Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central problem remains largely unsolved: how do we recover the finer texture details when we super-resolve at large upscaling factors? The behavior of optimization-based super-resolution methods is principally driven by the choice of the objective function. Recent work has largely focused on minimizing the mean squared reconstruction error. The resulting estimates have high peak signal-to-noise ratios, but they are often lacking high-frequency details and are perceptually unsatisfying in the sense that they fail to match the fidelity expected at the higher resolution. In this paper, we present SRGAN, a generative adversarial network (GAN) for image super-resolution (SR). To our knowledge, it is the first framework capable of inferring photo-realistic natural images for 4x upscaling factors. To achieve this, we propose a perceptual loss function which consists of an adversarial loss and a content loss. The adversarial loss pushes our solution to the natural image manifold using a discriminator network that is trained to differentiate between the super-resolved images and original photo-realistic images. In addition, we use a content loss motivated by perceptual similarity instead of similarity in pixel space. Our deep residual network is able to recover photo-realistic textures from heavily downsampled images on public benchmarks. An extensive mean-opinion-score (MOS) test shows hugely significant gains in perceptual quality using SRGAN. The MOS scores obtained with SRGAN are closer to those of the original high-resolution images than to those obtained with any state-of-the-art method.

/pdf/photo-realistic-single-image-super-resolution-using-a-2y0438rlmd.pdf

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

We consider image transformation problems, where an input image is transformed into an output image. Recent methods for such problems typically train feed-forward convolutional neural networks using a per-pixel loss between the output and ground-truth images. Parallel work has shown that high-quality images can be generated by defining and optimizing perceptual loss functions based on high-level features extracted from pretrained networks. We combine the benefits of both approaches, and propose the use of perceptual loss functions for training feed-forward networks for image transformation tasks. We show results on image style transfer, where a feed-forward network is trained to solve the optimization problem proposed by Gatys et al. in real-time. Compared to the optimization-based method, our network gives similar qualitative results but is three orders of magnitude faster. We also experiment with single-image super-resolution, where replacing a per-pixel loss with a perceptual loss gives visually pleasing results.

/pdf/perceptual-losses-for-real-time-style-transfer-and-super-hq4q54uf80.pdf

Perceptual Losses for Real-Time Style Transfer and Super-Resolution

We propose a deep learning method for single image super-resolution (SR). Our method directly learns an end-to-end mapping between the low/high-resolution images. The mapping is represented as a deep convolutional neural network (CNN) that takes the low-resolution image as the input and outputs the high-resolution one. We further show that traditional sparse-coding-based SR methods can also be viewed as a deep convolutional network. But unlike traditional methods that handle each component separately, our method jointly optimizes all layers. Our deep CNN has a lightweight structure, yet demonstrates state-of-the-art restoration quality, and achieves fast speed for practical on-line usage. We explore different network structures and parameter settings to achieve trade-offs between performance and speed. Moreover, we extend our network to cope with three color channels simultaneously, and show better overall reconstruction quality.

Image Super-Resolution Using Deep Convolutional Networks

We consider image transformation problems, where an input image is transformed into an output image. Recent methods for such problems typically train feed-forward convolutional neural networks using a \emph{per-pixel} loss between the output and ground-truth images. Parallel work has shown that high-quality images can be generated by defining and optimizing \emph{perceptual} loss functions based on high-level features extracted from pretrained networks. We combine the benefits of both approaches, and propose the use of perceptual loss functions for training feed-forward networks for image transformation tasks. We show results on image style transfer, where a feed-forward network is trained to solve the optimization problem proposed by Gatys et al in real-time. Compared to the optimization-based method, our network gives similar qualitative results but is three orders of magnitude faster. We also experiment with single-image super-resolution, where replacing a per-pixel loss with a perceptual loss gives visually pleasing results.

Hyperspectral recovery from a single RGB image has seen a great improvement with the development of deep convolutional neural networks (CNNs). In this paper, we propose two advanced CNNs for the hyperspectral reconstruction task, collectively called HSCNN+. We first develop a deep residual network named HSCNN-R, which comprises a number of residual blocks. The superior performance of this model comes from the modern architecture and optimization by removing the hand-crafted upsampling in HSCNN. Based on the promising results of HSCNN-R, we propose another distinct architecture that replaces the residual block by the dense block with a novel fusion scheme, leading to a new network named HSCNN-D. This model substantially deepens the network structure for a more accurate solution. Experimental results demonstrate that our proposed models significantly advance the state-of-the-art. In the NTIRE 2018 Spectral Reconstruction Challenge, our entries rank the 1st (HSCNN-D) and 2nd (HSCNN-R) places on both the "Clean" and "Real World" tracks. (Codes are available at [clean-r], [realworld-r], [clean-d], and [realworld-d].)

/pdf/hscnn-advanced-cnn-based-hyperspectral-recovery-from-rgb-1tdvyp8m3f.pdf

HSCNN+: Advanced CNN-Based Hyperspectral Recovery from RGB Images

We propose a pseudo-sequence-based scheme for light field image compression. In our scheme, the raw image captured by a light field camera is decomposed into multiple views according to the lenslet array of that camera. These views constitute a pseudo sequence like video, and the redundancy between views is exploited by a video encoder. The specific coding order of views, prediction structure, and rate allocation have been investigated for encoding the pseudo sequence. Experimental results show the superior performance of our scheme, which achieves as high as 6.6 dB gain compared with directly encoding the raw image by the legacy JPEG.

Pseudo-sequence-based light field image compression

The new H.264 (MPEG-4 AVC) video coding standard can achieve considerably higher coding efficiency compared to previous standards. This is accomplished mainly due to the consideration of variable block sizes for motion compensation, multiple reference frames, intra prediction, but also due to better exploitation of the spatiotemporal correlation that may exist between adjacent Macroblocks, with the SKIP mode in predictive (P) slices and the two DIRECT modes in bipredictive (B) slices. These modes, when signaled, could in effect represent the motion of a macroblock (MB) or block without having to transmit any additional motion information required by other inter-MB types. This property also allows these modes to be highly compressible especially due to the consideration of run length coding strategies. Although spatial correlation of motion vectors from adjacent MBs is used for SKIP mode to predict its motion parameters, until recently, DIRECT mode considered only temporal correlation of adjacent pictures. In this letter, we introduce alternative methods for the generation of the motion information for the DIRECT mode using spatial or combined spatiotemporal correlation. Considering that temporal correlation requires that the motion and timestamp information from previous pictures are available in both the encoder and decoder, it is shown that our spatial-only method can reduce or eliminate such requirements while, at the same time, achieving similar performance. The combined methods, on the other hand, by jointly exploiting spatial and temporal correlation either at the MB or slice/picture level, can achieve even higher coding efficiency. Finally, improvements on the existing Rate Distortion Optimization related to B slices within the H.264 codec are also presented, which can lead to improvements of up to 16% in bit rate reduction or, equivalently, more than 0.7 dB in PSNR.

Direct mode coding for bipredictive slices in the H.264 standard

This paper presents a unified deep learning framework to recover hyperspectral images from spectrally undersampled projections. Specifically, we investigate two kinds of representative projections, RGB and compressive sensing (CS) measurements. These measurements are first upsampled in the spectral dimension through simple interpolation or CS reconstruction, and the proposed method learns an end-to-end mapping from a large number of up-sampled/groundtruth hyperspectral image pairs. The mapping is represented as a deep convolutional neural network (CNN) that takes the spectrally upsampled image as input and outputs the enhanced hyperspetral one. We explore different network configurations to achieve high reconstruction fidelity. Experimental results on a variety of test images demonstrate significantly improved performance of the proposed method over the state-of-the-arts.

/pdf/hscnn-cnn-based-hyperspectral-image-recovery-from-spectrally-15kk17rug8.pdf

HSCNN: CNN-Based Hyperspectral Image Recovery from Spectrally Undersampled Projections

A motion-compensated video encoding scheme employs progressive fine-granularity layered coding to encode macroblocks of video data into frames having multiple layers, including a base layer of comparatively low quality video and multiple enhancement layers of increasingly higher quality video. Some of the enhancement layers in a current frame are predicted from different quality layers in reference frames. The video encoding scheme estimates drifting errors during the encoding and chooses a coding mode for each macroblock in the enhancement layer to maximize high coding efficiency while minimizing drifting errors.

Feng Wu

Papers

HSCNN+: Advanced CNN-Based Hyperspectral Recovery from RGB Images

Pseudo-sequence-based light field image compression

Direct mode coding for bipredictive slices in the H.264 standard

HSCNN: CNN-Based Hyperspectral Image Recovery from Spectrally Undersampled Projections

Drifting reduction and macroblock-based control in progressive fine granularity scalable video coding