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 superresolution (SR). To our knowledge, it is the first framework capable of inferring photo-realistic natural images for 4× 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. arXiv:1609.04802v5 [cs.CV]

Recently, several models based on deep neural networks have achieved great success in terms of both reconstruction accuracy and computational performance for single image super-resolution. In these methods, the low resolution (LR) input image is upscaled to the high resolution (HR) space using a single filter, commonly bicubic interpolation, before reconstruction. This means that the super-resolution (SR) operation is performed in HR space. We demonstrate that this is sub-optimal and adds computational complexity. In this paper, we present the first convolutional neural network (CNN) capable of real-time SR of 1080p videos on a single K2 GPU. To achieve this, we propose a novel CNN architecture where the feature maps are extracted in the LR space. In addition, we introduce an efficient sub-pixel convolution layer which learns an array of upscaling filters to upscale the final LR feature maps into the HR output. By doing so, we effectively replace the handcrafted bicubic filter in the SR pipeline with more complex upscaling filters specifically trained for each feature map, whilst also reducing the computational complexity of the overall SR operation. We evaluate the proposed approach using images and videos from publicly available datasets and show that it performs significantly better (+0.15dB on Images and +0.39dB on Videos) and is an order of magnitude faster than previous CNN-based methods.

Convolutional neural networks have enabled accurate image super-resolution in real-time. However, recent attempts to benefit from temporal correlations in video superresolution have been limited to naive or inefficient architectures. In this paper, we introduce spatio-temporal subpixel convolution networks that effectively exploit temporal redundancies and improve reconstruction accuracy while maintaining real-time speed. Specifically, we discuss the use of early fusion, slow fusion and 3D convolutions for the joint processing of multiple consecutive video frames. We also propose a novel joint motion compensation and video super-resolution algorithm that is orders of magnitude more efficient than competing methods, relying on a fast multi-resolution spatial transformer module that is endto-end trainable. These contributions provide both higher accuracy and temporally more consistent videos, which we confirm qualitatively and quantitatively. Relative to singleframe models, spatio-temporal networks can either reduce the computational cost by 30% whilst maintaining the same quality or provide a 0.2dB gain for a similar computational cost. Results on publicly available datasets demonstrate that the proposed algorithms surpass current state-of-theart performance in both accuracy and efficiency.

No abstract

We propose a new approach to the problem of optimizing autoencoders for lossy image compression. New media formats, changing hardware technology, as well as diverse requirements and content types create a need for compression algorithms which are more flexible than existing codecs. Autoencoders have the potential to address this need, but are difficult to optimize directly due to the inherent non-differentiabilty of the compression loss. We here show that minimal changes to the loss are sufficient to train deep autoencoders competitive with JPEG 2000 and outperforming recently proposed approaches based on RNNs. Our network is furthermore computationally efficient thanks to a sub-pixel architecture, which makes it suitable for high-resolution images. This is in contrast to previous work on autoencoders for compression using coarser approximations, shallower architectures, computationally expensive methods, or focusing on small images.

BackgroundLate Gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) imaging can be used to visualise regions of fibrosis and scarring in the left atrium (LA) myocardium. This can be important for treatment stratification of patients with atrial fibrillation (AF) and for assessment of treatment after radio frequency catheter ablation (RFCA). In this paper we present a standardised evaluation benchmarking framework for algorithms segmenting fibrosis and scar from LGE CMR images. The algorithms reported are the response to an open challenge that was put to the medical imaging community through an ISBI (IEEE International Symposium on Biomedical Imaging) workshop.MethodsThe image database consisted of 60 multicenter, multivendor LGE CMR image datasets from patients with AF, with 30 images taken before and 30 after RFCA for the treatment of AF. A reference standard for scar and fibrosis was established by merging manual segmentations from three observers. Furthermore, scar was also quantified using 2, 3 and 4 standard deviations (SD) and full-width-at-half-maximum (FWHM) methods. Seven institutions responded to the challenge: Imperial College (IC), Mevis Fraunhofer (MV), Sunnybrook Health Sciences (SY), Harvard/Boston University (HB), Yale School of Medicine (YL), King’s College London (KCL) and Utah CARMA (UTA, UTB). There were 8 different algorithms evaluated in this study.ResultsSome algorithms were able to perform significantly better than SD and FWHM methods in both pre- and post-ablation imaging. Segmentation in pre-ablation images was challenging and good correlation with the reference standard was found in post-ablation images. Overlap scores (out of 100) with the reference standard were as follows: Pre: IC = 37, MV = 22, SY = 17, YL = 48, KCL = 30, UTA = 42, UTB = 45; Post: IC = 76, MV = 85, SY = 73, HB = 76, YL = 84, KCL = 78, UTA = 78, UTB = 72.ConclusionsThe study concludes that currently no algorithm is deemed clearly better than others. There is scope for further algorithmic developments in LA fibrosis and scar quantification from LGE CMR images. Benchmarking of future scar segmentation algorithms is thus important. The proposed benchmarking framework is made available as open-source and new participants can evaluate their algorithms via a web-based interface.

We consider the problem of face swapping in images, where an input identity is transformed into a target identity while preserving pose, facial expression and lighting. To perform this mapping, we use convolutional neural networks trained to capture the appearance of the target identity from an unstructured collection of his/her photographs. This approach is enabled by framing the face swapping problem in terms of style transfer, where the goal is to render an image in the style of another one. Building on recent advances in this area, we devise a new loss function that enables the network to produce highly photorealistic results. By combining neural networks with simple pre-and post-processing steps, we aim at making face swap work in real-time with no input from the user.

In this paper we present a benchmarking framework for the validation of cardiac motion analysis algorithms. The reported methods are the response to an open challenge that was put to the medical imaging community through a MICCAI workshop. The database included magnetic resonance (MR) and 3D ultrasound (3DUS) datasets from a dynamic phantom and 15 healthy volunteers. Participants processed 3D tagged MR datasets (3DTAG), cine steady state free precession MR datasets (SSFP) and 3DUS datasets, amounting to 1158 image volumes. Ground-truth for motion tracking was based on 12 landmarks (4 walls at 3 ventricular levels). They were manually tracked by two observers in the 3DTAG data over the whole cardiac cycle, using an in-house application with 4D visualization capabilities. The median of the inter-observer variability was computed for the phantom dataset (0.77mm) and for the volunteer datasets (0.84mm). The ground-truth was registered to 3DUS coordinates using a point based similarity transform. Four institutions responded to the challenge by providing motion estimates for the data: Fraunhofer MEVIS (MEVIS), Bremen, Germany; Imperial College London -University College London (IUCL), UK; Universitat Pompeu Fabra (UPF), Barcelona, Spain; Inria-Asclepios project (INRIA), France. Details on the implementation and evaluation of the four methodologies are presented in this manuscript. The manually tracked landmarks were used to evaluate tracking accuracy of all methodologies. For 3DTAG, median values were computed over all time frames for the phantom dataset (MEVIS=1.20mm, IUCL=0.73mm, UPF=1.10mm, INRIA=1.09mm) and for the volunteer datasets (MEVIS=1.33mm, IUCL=1.52mm, UPF=1.09mm, INRIA=1.32mm). For 3DUS, median values were computed at end diastole and end systole for the phantom dataset (MEVIS=4.40mm, UPF=3.48mm, INRIA=4.78mm) and for the volunteer datasets (MEVIS=3.51mm, UPF=3.71mm, INRIA=4.07mm). For SSFP, median values were computed at end diastole and end systole for the phantom dataset (UPF=6.18mm, INRIA=3.93mm) and for the volunteer datasets (UPF=3.09mm, INRIA=4.78mm). Finally, strain curves were generated and qualitatively compared. Good agreement was found between the different modalities and methodologies, except for radial strain that showed a high variability in cases of lower image quality.