Deep learning-based super-resolution (SR) techniques have generally achieved excellent performance in the computer vision field. Recently, it has been proven that three-dimensional (3D) SR for medical volumetric data delivers better visual results than conventional two-dimensional (2D) processing. However, deepening and widening 3D networks increases training difficulty significantly due to the large number of parameters and small number of training samples. Thus, we propose a 3D convolutional neural network (CNN) for SR of medical volumetric data called ParallelNet using parallel connections. We construct a parallel connection structure based on the group convolution and feature aggregation to build a 3D CNN that is as wide as possible with few parameters. As a result, the model thoroughly learns more feature maps with larger receptive fields. In addition, to further improve accuracy, we present an efficient version of ParallelNet (called VolumeNet), which reduces the number of parameters and deepens ParallelNet using a proposed lightweight building block module called the Queue module. Unlike most lightweight CNNs based on depthwise convolutions, the Queue module is primarily constructed using separable 2D cross-channel convolutions. As a result, the number of network parameters and computational complexity can be reduced significantly while maintaining accuracy due to full channel fusion. Experimental results demonstrate that the proposed VolumeNet significantly reduces the number of model parameters and achieves high precision results compared to state-of-the-art methods.
Combining satellite imagery with machine learning (SIML) has the potential to address global challenges by remotely estimating socioeconomic and environmental conditions in data-poor regions, yet the resource requirements of SIML limit its accessibility and use. We show that a single encoding of satellite imagery can generalize across diverse prediction tasks (e.g. forest cover, house price, road length). Our method achieves accuracy competitive with deep neural networks at orders of magnitude lower computational cost, scales globally, delivers label super-resolution predictions, and facilitates characterizations of uncertainty. Since image encodings are shared across tasks, they can be centrally computed and distributed to unlimited researchers, who need only fit a linear regression to their own ground truth data in order to achieve state-of-the-art SIML performance.
Cosmological simulations of galaxy formation are limited by finite computational resources. We draw from the ongoing rapid advances in Artificial Intelligence (specifically Deep Learning) to address this problem. Neural networks have been developed to learn from high-resolution (HR) image data, and then make accurate super-resolution (SR) versions of different low-resolution (LR) images. We apply such techniques to LR cosmological N-body simulations, generating SR versions. Specifically, we are able to enhance the simulation resolution by generating 512 times more particles and predicting their displacements from the initial positions. Therefore our results can be viewed as new simulation realizations themselves rather than projections, e.g., to their density fields. Furthermore, the generation process is stochastic, enabling us to sample the small-scale modes conditioning on the large-scale environment. Our model learns from only 16 pairs of small-volume LR-HR simulations, and is then able to generate SR simulations that successfully reproduce the HR matter power spectrum to percent level up to $16\,h^{-1}\mathrm{Mpc}$, and the HR halo mass function to within $10 \%$ down to $10^{11} \, M_\odot$. We successfully deploy the model in a box 1000 times larger than the training simulation box, showing that high-resolution mock surveys can be generated rapidly. We conclude that AI assistance has the potential to revolutionize modeling of small-scale galaxy formation physics in large cosmological volumes.
The collection of high-resolution training data is crucial in building robust plant disease diagnosis systems, since such data have a significant impact on diagnostic performance. However, they are very difficult to obtain and are not always available in practice. Deep learning-based techniques, and particularly generative adversarial networks (GANs), can be applied to generate high-quality super-resolution images, but these methods often produce unexpected artifacts that can lower the diagnostic performance. In this paper, we propose a novel artifact-suppression super-resolution method that is specifically designed for diagnosing leaf disease, called Leaf Artifact-Suppression Super Resolution (LASSR). Thanks to its own artifact removal module that detects and suppresses artifacts to a considerable extent, LASSR can generate much more pleasing, high-quality images compared to the state-of-the-art ESRGAN model. Experiments based on a five-class cucumber disease (including healthy) discrimination model show that training with data generated by LASSR significantly boosts the performance on an unseen test dataset by nearly 22% compared with the baseline, and that our approach is more than 2% better than a model trained with images generated by ESRGAN.
Recurrent models are a popular choice for video enhancement tasks such as video denoising or super-resolution. In this work, we focus on their stability as dynamical systems and show that they tend to fail catastrophically at inference time on long video sequences. To address this issue, we (1) introduce a diagnostic tool which produces input sequences optimized to trigger instabilities and that can be interpreted as visualizations of temporal receptive fields, and (2) propose two approaches to enforce the stability of a model during training: constraining the spectral norm or constraining the stable rank of its convolutional layers. We then introduce Stable Rank Normalization for Convolutional layers (SRN-C), a new algorithm that enforces these constraints. Our experimental results suggest that SRN-C successfully enforces stability in recurrent video processing models without a significant performance loss.
Image-to-Image (I2I) translation is a heated topic in academia, and it also has been applied in real-world industry for tasks like image synthesis, super-resolution, and colorization. However, traditional I2I translation methods train data in two or more domains together. This requires lots of computation resources. Moreover, the results are of lower quality, and they contain many more artifacts. The training process could be unstable when the data in different domains are not balanced, and modal collapse is more likely to happen. We proposed a new I2I translation method that generates a new model in the target domain via a series of model transformations on a pre-trained StyleGAN2 model in the source domain. After that, we proposed an inversion method to achieve the conversion between an image and its latent vector. By feeding the latent vector into the generated model, we can perform I2I translation between the source domain and target domain. Both qualitative and quantitative evaluations were conducted to prove that the proposed method can achieve outstanding performance in terms of image quality, diversity and semantic similarity to the input and reference images compared to state-of-the-art works.
Internet-enabled smartphones and ultra-wide displays are transforming a variety of visual apps spanning from on-demand movies and 360-degree videos to video-conferencing and live streaming. However, robustly delivering visual content under fluctuating networking conditions on devices of diverse capabilities remains an open problem. In recent years, advances in the field of deep learning on tasks such as super-resolution and image enhancement have led to unprecedented performance in generating high-quality images from low-quality ones, a process we refer to as neural enhancement. In this paper, we survey state-of-the-art content delivery systems that employ neural enhancement as a key component in achieving both fast response time and high visual quality. We first present the deployment challenges of neural enhancement models. We then cover systems targeting diverse use-cases and analyze their design decisions in overcoming technical challenges. Moreover, we present promising directions based on the latest insights from deep learning research to further boost the quality of experience of these systems.
In the last few years, several deep learning models, especially Generative Adversarial Networks have received a lot of attention for the task of Single Image Super-Resolution (SISR). These methods focus on building an end-to-end framework, which produce a high resolution(SR) image from a given low resolution(LR) image in a single step to achieve state-of-the-art performance. This paper focuses on improving an existing deep-learning based method to perform Super-Resolution Microscopy in real-time using a standard GPU. For this, we first propose a tiling strategy, which takes advantage of parallelism provided by a GPU to speed up the network training process. Further, we suggest simple changes to the architecture of the generator and the discriminator of SRGAN. Subsequently, We compare the quality and the running time for the outputs produced by our model, opening its applications in different areas like low-end benchtop and even mobile microscopy. Finally, we explore the possibility of the trained network to produce High-Resolution HR outputs for different domains.
Deep convolutional neural networks have significantly improved the peak signal-to-noise ratio of SuperResolution (SR). However, image viewer applications commonly allow users to zoom the images to arbitrary magnification scales, thus far imposing a large number of required training scales at a tremendous computational cost. To obtain a more computationally efficient model for arbitrary scale SR, this paper employs a Laplacian pyramid method to reconstruct any-scale high-resolution (HR) images using the high-frequency image details in a Laplacian Frequency Representation. For SR of small-scales (between 1 and 2), images are constructed by interpolation from a sparse set of precalculated Laplacian pyramid levels. SR of larger scales is computed by recursion from small scales, which significantly reduces the computational cost. For a full comparison, fixed- and any-scale experiments are conducted using various benchmarks. At fixed scales, ASDN outperforms predefined upsampling methods (e.g., SRCNN, VDSR, DRRN) by about 1 dB in PSNR. At any-scale, ASDN generally exceeds Meta-SR on many scales.
Divide and conquer is an established algorithm design paradigm that has proven itself to solve a variety of problems efficiently. However, it is yet to be fully explored in solving problems with a neural network, particularly the problem of image super-resolution. In this work, we propose an approach to divide the problem of image super-resolution into multiple sub-problems and then solve/conquer them with the help of a neural network. Unlike a typical deep neural network, we design an alternate network architecture that is much wider (along with being deeper) than existing networks and is specially designed to implement the divide-and-conquer design paradigm with a neural network. Additionally, a technique to calibrate the intensities of feature map pixels is being introduced. Extensive experimentation on five datasets reveals that our approach towards the problem and the proposed architecture generate better and sharper results than current state-of-the-art methods.
Previous methods decompose blind super resolution (SR) problem into two sequential steps: \textit{i}) estimating blur kernel from given low-resolution (LR) image and \textit{ii}) restoring SR image based on estimated kernel. This two-step solution involves two independently trained models, which may not be well compatible with each other. Small estimation error of the first step could cause severe performance drop of the second one. While on the other hand, the first step can only utilize limited information from LR image, which makes it difficult to predict highly accurate blur kernel. Towards these issues, instead of considering these two steps separately, we adopt an alternating optimization algorithm, which can estimate blur kernel and restore SR image in a single model. Specifically, we design two convolutional neural modules, namely \textit{Restorer} and \textit{Estimator}. \textit{Restorer} restores SR image based on predicted kernel, and \textit{Estimator} estimates blur kernel with the help of restored SR image. We alternate these two modules repeatedly and unfold this process to form an end-to-end trainable network. In this way, \textit{Estimator} utilizes information from both LR and SR images, which makes the estimation of blur kernel easier. More importantly, \textit{Restorer} is trained with the kernel estimated by \textit{Estimator}, instead of ground-truth kernel, thus \textit{Restorer} could be more tolerant to the estimation error of \textit{Estimator}. Extensive experiments on synthetic datasets and real-world images show that our model can largely outperform state-of-the-art methods and produce more visually favorable results at much higher speed. The source code is available at https://github.com/greatlog/DAN.git.
As the success of Generative Adversarial Networks (GANs) on natural images quickly propels them into various real-life applications across different domains, it becomes more and more important to clearly understand their limitations. Specifically, understanding GANs' capability across the full spectrum of spatial frequencies, i.e. beyond the low-frequency dominant spectrum of natural images, is critical for assessing the reliability of GAN generated data in any detail-sensitive application (e.g. denoising, filling and super-resolution in medical and satellite images). In this paper, we show that the ability of convolutional GANs to learn a distribution is significantly affected by the spatial frequency of the underlying carrier signal, that is, GANs have a bias against learning high spatial frequencies. Crucially, we show that this bias is not merely a result of the scarcity of high frequencies in natural images, rather, it is a systemic bias hindering the learning of high frequencies regardless of their prominence in a dataset. Furthermore, we explain why large-scale GANs' ability to generate fine details on natural images does not exclude them from the adverse effects of this bias. Finally, we propose a method for manipulating this bias with minimal computational overhead. This method can be used to explicitly direct computational resources towards any specific spatial frequency of interest in a dataset, extending the flexibility of GANs.
Terrain, representing features of an earth surface, plays a crucial role in many applications such as simulations, route planning, analysis of surface dynamics, computer graphics-based games, entertainment, films, to name a few. With recent advancements in digital technology, these applications demand the presence of high-resolution details in the terrain. In this paper, we propose a novel fully convolutional neural network-based super-resolution architecture to increase the resolution of low-resolution Digital Elevation Model (LRDEM) with the help of information extracted from the corresponding aerial image as a complementary modality. We perform the super-resolution of LRDEM using an attention-based feedback mechanism named 'Attentional Feedback Network' (AFN), which selectively fuses the information from LRDEM and aerial image to enhance and infuse the high-frequency features and to produce the terrain realistically. We compare the proposed architecture with existing state-of-the-art DEM super-resolution methods and show that the proposed architecture outperforms enhancing the resolution of input LRDEM accurately and in a realistic manner.
Hyperspectral super-resolution refers to the problem of fusing a hyperspectral image (HSI) and a multispectral image (MSI) to produce a super-resolution image (SRI) that has fine spatial and spectral resolution. State-of-the-art methods approach the problem via low-rank matrix approximations to the matricized HSI and MSI. These methods are effective to some extent, but a number of challenges remain. First, HSIs and MSIs are naturally third-order tensors (data "cubes") and thus matricization is prone to loss of structural information--which could degrade performance. Second, it is unclear whether or not these low-rank matrix-based fusion strategies can guarantee identifiability or exact recovery of the SRI. However, identifiability plays a pivotal role in estimation problems and usually has a significant impact on performance in practice. Third, the majority of the existing methods assume that there are known (or easily estimated) degradation operators applied to the SRI to form the corresponding HSI and MSI--which is hardly the case in practice. In this work, we propose to tackle the super-resolution problem from a tensor perspective. Specifically, we utilize the multidimensional structure of the HSI and MSI to propose a coupled tensor factorization framework that can effectively overcome the aforementioned issues. The proposed approach guarantees the identifiability of the SRI under mild and realistic conditions. Furthermore, it works with little knowledge of the degradation operators, which is clearly an advantage over the existing methods. Semi-real numerical experiments are included to show the effectiveness of the proposed approach.
This work aims at designing a lightweight convolutional neural network for image super resolution (SR). With simplicity bare in mind, we construct a pretty concise and effective network with a newly proposed pixel attention scheme. Pixel attention (PA) is similar as channel attention and spatial attention in formulation. The difference is that PA produces 3D attention maps instead of a 1D attention vector or a 2D map. This attention scheme introduces fewer additional parameters but generates better SR results. On the basis of PA, we propose two building blocks for the main branch and the reconstruction branch, respectively. The first one - SC-PA block has the same structure as the Self-Calibrated convolution but with our PA layer. This block is much more efficient than conventional residual/dense blocks, for its twobranch architecture and attention scheme. While the second one - UPA block combines the nearest-neighbor upsampling, convolution and PA layers. It improves the final reconstruction quality with little parameter cost. Our final model- PAN could achieve similar performance as the lightweight networks - SRResNet and CARN, but with only 272K parameters (17.92% of SRResNet and 17.09% of CARN). The effectiveness of each proposed component is also validated by ablation study. The code is available at https://github.com/zhaohengyuan1/PAN.
High-quality magnetic resonance (MR) image, i.e., with near isotropic voxel spacing, is desirable in various scenarios of medical image analysis. However, many MR acquisitions use large inter-slice spacing in clinical practice. In this work, we propose a novel deep-learning-based super-resolution algorithm to generate high-resolution (HR) MR images with small slice spacing from low-resolution (LR) inputs of large slice spacing. Notice that most existing deep-learning-based methods need paired LR and HR images to supervise the training, but in clinical scenarios, usually no HR images will be acquired. Therefore, our unique goal herein is to design and train the super-resolution network with no real HR ground-truth. Specifically, two training stages are used in our method. First, HR images of reduced slice spacing are synthesized from real LR images using variational auto-encoder (VAE). Although these synthesized HR images are as realistic as possible, they may still suffer from unexpected morphing induced by VAE, implying that the synthesized HR images cannot be paired with the real LR images in terms of anatomical structure details. In the second stage, we degrade the synthesized HR images to generate corresponding LR images and train a super-resolution network based on these synthesized HR and degraded LR pairs. The underlying mechanism is that such a super-resolution network is less vulnerable to anatomical variability. Experiments on knee MR images successfully demonstrate the effectiveness of our proposed solution to reduce the slice spacing for better rendering.
Single image super-resolution is an effective way to enhance the spatial resolution of remote sensing image, which is crucial for many applications such as target detection and image classification. However, existing methods based on the neural network usually have small receptive fields and ignore the image detail. We propose a novel method named deep memory connected network (DMCN) based on a convolutional neural network to reconstruct high-quality super-resolution images. We build local and global memory connections to combine image detail with environmental information. To further reduce parameters and ease time-consuming, we propose downsampling units, shrinking the spatial size of feature maps. We test DMCN on three remote sensing datasets with different spatial resolution. Experimental results indicate that our method yields promising improvements in both accuracy and visual performance over the current state-of-the-art.
Video super-resolution, which attempts to reconstruct high-resolution video frames from their corresponding low-resolution versions, has received increasingly more attention in recent years. Most existing approaches opt to use deformable convolution to temporally align neighboring frames and apply traditional spatial attention mechanism (convolution based) to enhance reconstructed features. However, such spatial-only strategies cannot fully utilize temporal dependency among video frames. In this paper, we propose a novel deep learning based VSR algorithm, named Deformable Kernel Spatial Attention Network (DKSAN). Thanks to newly designed Deformable Kernel Convolution Alignment (DKC_Align) and Deformable Kernel Spatial Attention (DKSA) modules, DKSAN can better exploit both spatial and temporal redundancies to facilitate the information propagation across different layers. We have tested DKSAN on AIM2020 Video Extreme Super-Resolution Challenge to super-resolve videos with a scale factor as large as 16. Experimental results demonstrate that our proposed DKSAN can achieve both better subjective and objective performance compared with the existing state-of-the-art EDVR on Vid3oC and IntVID datasets.
Single image super-resolution (SISR) aims to recover the high-resolution (HR) image from its low-resolution (LR) input image. With the development of deep learning, SISR has achieved great progress. However, It is still a challenge to restore the real-world LR image with complicated authentic degradations. Therefore, we propose FAN, a frequency aggregation network, to address the real-world image super-resolu-tion problem. Specifically, we extract different frequencies of the LR image and pass them to a channel attention-grouped residual dense network (CA-GRDB) individually to output corresponding feature maps. And then aggregating these residual dense feature maps adaptively to recover the HR image with enhanced details and textures. We conduct extensive experiments quantitatively and qualitatively to verify that our FAN performs well on the real image super-resolution task of AIM 2020 challenge. According to the released final results, our team SR-IM achieves the fourth place on the X4 track with PSNR of 31.1735 and SSIM of 0.8728.
Light field (LF) images acquired by hand-held devices usually suffer from low spatial resolution as the limited detector resolution has to be shared with the angular dimension. LF spatial super-resolution (SR) thus becomes an indispensable part of the LF camera processing pipeline. The high-dimensionality characteristic and complex geometrical structure of LF images make the problem more challenging than traditional single-image SR. The performance of existing methods is still limited as they fail to thoroughly explore the coherence among LF sub-aperture images (SAIs) and are insufficient in accurately preserving the scene's parallax structure. To tackle this challenge, we propose a novel learning-based LF spatial SR framework. Specifically, each SAI of an LF image is first coarsely and individually super-resolved by exploring the complementary information among SAIs with selective combinatorial geometry embedding. To achieve efficient and effective selection of the complementary information, we propose two novel sub-modules conducted hierarchically: the patch selector provides an option of retrieving similar image patches based on offline disparity estimation to handle large-disparity correlations; and the SAI selector adaptively and flexibly selects the most informative SAIs to improve the embedding efficiency. To preserve the parallax structure among the reconstructed SAIs, we subsequently append a consistency regularization network trained over a structure-aware loss function to refine the parallax relationships over the coarse estimation. In addition, we extend the proposed method to irregular LF data. To the best of our knowledge, this is the first learning-based SR method for irregular LF data. Experimental results over both synthetic and real-world LF datasets demonstrate the significant advantage of our approach over state-of-the-art methods.