Normalizing flows have recently demonstrated promising results for low-level vision tasks. For image super-resolution (SR), it learns to predict diverse photo-realistic high-resolution (HR) images from the low-resolution (LR) image rather than learning a deterministic mapping. For image rescaling, it achieves high accuracy by jointly modelling the downscaling and upscaling processes. While existing approaches employ specialized techniques for these two tasks, we set out to unify them in a single formulation. In this paper, we propose the hierarchical conditional flow (HCFlow) as a unified framework for image SR and image rescaling. More specifically, HCFlow learns a bijective mapping between HR and LR image pairs by modelling the distribution of the LR image and the rest high-frequency component simultaneously. In particular, the high-frequency component is conditional on the LR image in a hierarchical manner. To further enhance the performance, other losses such as perceptual loss and GAN loss are combined with the commonly used negative log-likelihood loss in training. Extensive experiments on general image SR, face image SR and image rescaling have demonstrated that the proposed HCFlow achieves state-of-the-art performance in terms of both quantitative metrics and visual quality.
As an image sensing instrument, light field images can supply extra angular information compared with monocular images and have facilitated a wide range of measurement applications. Light field image capturing devices usually suffer from the inherent trade-off between the angular and spatial resolutions. To tackle this problem, several methods, such as light field reconstruction and light field super-resolution, have been proposed but leaving two problems unaddressed, namely domain asymmetry and efficient information flow. In this paper, we propose an end-to-end Spatio-Angular Dense Network (SADenseNet) for light field reconstruction with two novel components, namely correlation blocks and spatio-angular dense skip connections to address them. The former performs effective modeling of the correlation information in a way that conforms with the domain asymmetry. And the latter consists of three kinds of connections enhancing the information flow within two domains. Extensive experiments on both real-world and synthetic datasets have been conducted to demonstrate that the proposed SADenseNet's state-of-the-art performance at significantly reduced costs in memory and computation. The qualitative results show that the reconstructed light field images are sharp with correct details and can serve as pre-processing to improve the accuracy of related measurement applications.
Thermal images model the long-infrared range of the electromagnetic spectrum and provide meaningful information even when there is no visible illumination. Yet, unlike imagery that represents radiation from the visible continuum, infrared images are inherently low-resolution due to hardware constraints. The restoration of thermal images is critical for applications that involve safety, search and rescue, and military operations. In this paper, we introduce a system to efficiently reconstruct thermal images. Specifically, we explore how to effectively attend to contrasting receptive fields (RFs) where increasing the RFs of a network can be computationally expensive. For this purpose, we introduce a deep attention to varying receptive fields network (AVRFN). We supply a gated convolutional layer with higher-order information extracted from disparate RFs, whereby an RF is parameterized by a dilation rate. In this way, the dilation rate can be tuned to use fewer parameters thus increasing the efficacy of AVRFN. Our experimental results show an improvement over the state of the art when compared against competing thermal image super-resolution methods.
It has been shown that equivariant convolution is very helpful for many types of computer vision tasks. Recently, the 2D filter parametrization technique plays an important role when designing equivariant convolutions. However, the current filter parametrization method still has its evident drawbacks, where the most critical one lies in the accuracy problem of filter representation. Against this issue, in this paper we modify the classical Fourier series expansion for 2D filters, and propose a new set of atomic basis functions for filter parametrization. The proposed filter parametrization method not only finely represents 2D filters with zero error when the filter is not rotated, but also substantially alleviates the fence-effect-caused quality degradation when the filter is rotated. Accordingly, we construct a new equivariant convolution method based on the proposed filter parametrization method, named F-Conv. We prove that the equivariance of the proposed F-Conv is exact in the continuous domain, which becomes approximate only after discretization. Extensive experiments show the superiority of the proposed method. Particularly, we adopt rotation equivariant convolution methods to image super-resolution task, and F-Conv evidently outperforms previous filter parametrization based method in this task, reflecting its intrinsic capability of faithfully preserving rotation symmetries in local image features.
Bicubic downscaling is a prevalent technique used to reduce the video storage burden or to accelerate the downstream processing speed. However, the inverse upscaling step is non-trivial, and the downscaled video may also deteriorate the performance of downstream tasks. In this paper, we propose a self-conditioned probabilistic framework for video rescaling to learn the paired downscaling and upscaling procedures simultaneously. During the training, we decrease the entropy of the information lost in the downscaling by maximizing its probability conditioned on the strong spatial-temporal prior information within the downscaled video. After optimization, the downscaled video by our framework preserves more meaningful information, which is beneficial for both the upscaling step and the downstream tasks, e.g., video action recognition task. We further extend the framework to a lossy video compression system, in which a gradient estimator for non-differential industrial lossy codecs is proposed for the end-to-end training of the whole system. Extensive experimental results demonstrate the superiority of our approach on video rescaling, video compression, and efficient action recognition tasks.
Though many attempts have been made in blind super-resolution to restore low-resolution images with unknown and complex degradations, they are still far from addressing general real-world degraded images. In this work, we extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data. Specifically, a high-order degradation modeling process is introduced to better simulate complex real-world degradations. We also consider the common ringing and overshoot artifacts in the synthesis process. In addition, we employ a U-Net discriminator with spectral normalization to increase discriminator capability and stabilize the training dynamics. Extensive comparisons have shown its superior visual performance than prior works on various real datasets. We also provide efficient implementations to synthesize training pairs on the fly.
Video super-resolution (VSR) technology excels in reconstructing low-quality video, avoiding unpleasant blur effect caused by interpolation-based algorithms. However, vast computation complexity and memory occupation hampers the edge of deplorability and the runtime inference in real-life applications, especially for large-scale VSR task. This paper explores the possibility of real-time VSR system and designs an efficient and generic VSR network, termed EGVSR. The proposed EGVSR is based on spatio-temporal adversarial learning for temporal coherence. In order to pursue faster VSR processing ability up to 4K resolution, this paper tries to choose lightweight network structure and efficient upsampling method to reduce the computation required by EGVSR network under the guarantee of high visual quality. Besides, we implement the batch normalization computation fusion, convolutional acceleration algorithm and other neural network acceleration techniques on the actual hardware platform to optimize the inference process of EGVSR network. Finally, our EGVSR achieves the real-time processing capacity of [email omitted]. Compared with TecoGAN, the most advanced VSR network at present, we achieve 85.04% reduction of computation density and 7.92x performance speedups. In terms of visual quality, the proposed EGVSR tops the list of most metrics (such as LPIPS, tOF, tLP, etc.) on the public test dataset Vid4 and surpasses other state-of-the-art methods in overall performance score. The source code of this project can be found on https://github.com/Thmen/EGVSR.
In fluorescence microscopy, Single Molecule Localization Microscopy (SMLM) techniques aim at localizing with high precision high density fluorescent molecules by stochastically activating and imaging small subsets of blinking emitters. Super Resolution (SR) plays an important role in this field since it allows to go beyond the intrinsic light diffraction limit. In this work, we propose a deep learning-based algorithm for precise molecule localization of high density frames acquired by SMLM techniques whose $\ell_{2}$-based loss function is regularized by positivity and $\ell_{0}$-based constraints. The $\ell_{0}$ is relaxed through its Continuous Exact $\ell_{0}$ (CEL0) counterpart. The arising approach, named DeepCEL0, is parameter-free, more flexible, faster and provides more precise molecule localization maps if compared to the other state-of-the-art methods. We validate our approach on both simulated and real fluorescence microscopy data.
Hyperspectral pansharpening aims to synthesize a low-resolution hyperspectral image (LR-HSI) with a registered panchromatic image (PAN) to generate an enhanced HSI with high spectral and spatial resolution. Recently proposed HS pansharpening methods have obtained remarkable results using deep convolutional networks (ConvNets), which typically consist of three steps: (1) up-sampling the LR-HSI, (2) predicting the residual image via a ConvNet, and (3) obtaining the final fused HSI by adding the outputs from first and second steps. Recent methods have leveraged Deep Image Prior (DIP) to up-sample the LR-HSI due to its excellent ability to preserve both spatial and spectral information, without learning from large data sets. However, we observed that the quality of up-sampled HSIs can be further improved by introducing an additional spatial-domain constraint to the conventional spectral-domain energy function. We define our spatial-domain constraint as the $L_1$ distance between the predicted PAN image and the actual PAN image. To estimate the PAN image of the up-sampled HSI, we also propose a learnable spectral response function (SRF). Moreover, we noticed that the residual image between the up-sampled HSI and the reference HSI mainly consists of edge information and very fine structures. In order to accurately estimate fine information, we propose a novel over-complete network, called HyperKite, which focuses on learning high-level features by constraining the receptive from increasing in the deep layers. We perform experiments on three HSI datasets to demonstrate the superiority of our DIP-HyperKite over the state-of-the-art pansharpening methods. The deployment codes, pre-trained models, and final fusion outputs of our DIP-HyperKite and the methods used for the comparisons will be publicly made available at https://github.com/wgcban/DIP-HyperKite.git.
The deep image prior showed that a randomly initialized network with a suitable architecture can be trained to solve inverse imaging problems by simply optimizing it's parameters to reconstruct a single degraded image. However, it suffers from two practical limitations. First, it remains unclear how to control the prior beyond the choice of the network architecture. Second, training requires an oracle stopping criterion as during the optimization the performance degrades after reaching an optimum value. To address these challenges we introduce a frequency-band correspondence measure to characterize the spectral bias of the deep image prior, where low-frequency image signals are learned faster and better than high-frequency counterparts. Based on our observations, we propose techniques to prevent the eventual performance degradation and accelerate convergence. We introduce a Lipschitz-controlled convolution layer and a Gaussian-controlled upsampling layer as plug-in replacements for layers used in the deep architectures. The experiments show that with these changes the performance does not degrade during optimization, relieving us from the need for an oracle stopping criterion. We further outline a stopping criterion to avoid superfluous computation. Finally, we show that our approach obtains favorable results compared to current approaches across various denoising, deblocking, inpainting, super-resolution and detail enhancement tasks. Code is available at \url{https://github.com/shizenglin/Measure-and-Control-Spectral-Bias}.
While researches on model-based blind single image super-resolution (SISR) have achieved tremendous successes recently, most of them do not consider the image degradation sufficiently. Firstly, they always assume image noise obeys an independent and identically distributed (i.i.d.) Gaussian or Laplacian distribution, which largely underestimates the complexity of real noise. Secondly, previous commonly-used kernel priors (e.g., normalization, sparsity) are not effective enough to guarantee a rational kernel solution, and thus degenerates the performance of subsequent SISR task. To address the above issues, this paper proposes a model-based blind SISR method under the probabilistic framework, which elaborately models image degradation from the perspectives of noise and blur kernel. Specifically, instead of the traditional i.i.d. noise assumption, a patch-based non-i.i.d. noise model is proposed to tackle the complicated real noise, expecting to increase the degrees of freedom of the model for noise representation. As for the blur kernel, we novelly construct a concise yet effective kernel generator, and plug it into the proposed blind SISR method as an explicit kernel prior (EKP). To solve the proposed model, a theoretically grounded Monte Carlo EM algorithm is specifically designed. Comprehensive experiments demonstrate the superiority of our method over current state-of-the-arts on synthetic and real datasets. The source code is available at https://github.com/zsyOAOA/BSRDM.
Scene text image super-resolution (STISR) aims to improve the resolution and visual quality of low-resolution (LR) scene text images, and consequently boost the performance of text recognition. However, most of existing STISR methods regard text images as natural scene images, ignoring the categorical information of text. In this paper, we make an inspiring attempt to embed categorical text prior into STISR model training. Specifically, we adopt the character probability sequence as the text prior, which can be obtained conveniently from a text recognition model. The text prior provides categorical guidance to recover high-resolution (HR) text images. On the other hand, the reconstructed HR image can refine the text prior in return. Finally, we present a multi-stage text prior guided super-resolution (TPGSR) framework for STISR. Our experiments on the benchmark TextZoom dataset show that TPGSR can not only effectively improve the visual quality of scene text images, but also significantly improve the text recognition accuracy over existing STISR methods. Our model trained on TextZoom also demonstrates certain generalization capability to the LR images in other datasets.
Accelerated multi-modal magnetic resonance (MR) imaging is a new and effective solution for fast MR imaging, providing superior performance in restoring the target modality from its undersampled counterpart with guidance from an auxiliary modality. However, existing works simply combine the auxiliary modality as prior information, lacking in-depth investigations on the potential mechanisms for fusing different modalities. Further, they usually rely on the convolutional neural networks (CNNs), which is limited by the intrinsic locality in capturing the long-distance dependency. To this end, we propose a multi-modal transformer (MTrans), which is capable of transferring multi-scale features from the target modality to the auxiliary modality, for accelerated MR imaging. To capture deep multi-modal information, our MTrans utilizes an improved multi-head attention mechanism, named cross attention module, which absorbs features from the auxiliary modality that contribute to the target modality. Our framework provides three appealing benefits: (i) Our MTrans use an improved transformers for multi-modal MR imaging, affording more global information compared with existing CNN-based methods. (ii) A new cross attention module is proposed to exploit the useful information in each modality at different scales. The small patch in the target modality aims to keep more fine details, the large patch in the auxiliary modality aims to obtain high-level context features from the larger region and supplement the target modality effectively. (iii) We evaluate MTrans with various accelerated multi-modal MR imaging tasks, e.g., MR image reconstruction and super-resolution, where MTrans outperforms state-of-the-art methods on fastMRI and real-world clinical datasets.
Recent supervised point cloud upsampling methods are restricted by the size of training data and are limited in terms of covering all object shapes. Besides the challenges faced due to data acquisition, the networks also struggle to generalize on unseen records. In this paper, we present an internal point cloud upsampling approach at a holistic level referred to as "Zero-Shot" Point Cloud Upsampling (ZSPU). Our approach is data agnostic and relies solely on the internal information provided by a particular point cloud without patching in both self-training and testing phases. This single-stream design significantly reduces the training time by learning the relation between low resolution (LR) point clouds and their high (original) resolution (HR) counterparts. This association will then provide super resolution (SR) outputs when original point clouds are loaded as input. ZSPU achieves competitive/superior quantitative and qualitative performances on benchmark datasets when compared with other upsampling methods.
Monte Carlo rendering algorithms are widely used to produce photorealistic computer graphics images. However, these algorithms need to sample a substantial amount of rays per pixel to enable proper global illumination and thus require an immense amount of computation. In this paper, we present a hybrid rendering method to speed up Monte Carlo rendering algorithms. Our method first generates two versions of a rendering: one at a low resolution with a high sample rate (LRHS) and the other at a high resolution with a low sample rate (HRLS). We then develop a deep convolutional neural network to fuse these two renderings into a high-quality image as if it were rendered at a high resolution with a high sample rate. Specifically, we formulate this fusion task as a super resolution problem that generates a high resolution rendering from a low resolution input (LRHS), assisted with the HRLS rendering. The HRLS rendering provides critical high frequency details which are difficult to recover from the LRHS for any super resolution methods. Our experiments show that our hybrid rendering algorithm is significantly faster than the state-of-the-art Monte Carlo denoising methods while rendering high-quality images when tested on both our own BCR dataset and the Gharbi dataset. \url{https://github.com/hqqxyy/msspl}
Normalizing flows are a powerful class of generative models demonstrating strong performance in several speech and vision problems. In contrast to other generative models, normalizing flows are latent variable models with tractable likelihoods and allow for stable training. However, they have to be carefully designed to represent invertible functions with efficient Jacobian determinant calculation. In practice, these requirements lead to overparameterized and sophisticated architectures that are inferior to alternative feed-forward models in terms of inference time and memory consumption. In this work, we investigate whether one can distill flow-based models into more efficient alternatives. We provide a positive answer to this question by proposing a simple distillation approach and demonstrating its effectiveness on state-of-the-art conditional flow-based models for image super-resolution and speech synthesis.
Fetal motion is unpredictable and rapid on the scale of conventional MR scan times. Therefore, dynamic fetal MRI, which aims at capturing fetal motion and dynamics of fetal function, is limited to fast imaging techniques with compromises in image quality and resolution. Super-resolution for dynamic fetal MRI is still a challenge, especially when multi-oriented stacks of image slices for oversampling are not available and high temporal resolution for recording the dynamics of the fetus or placenta is desired. Further, fetal motion makes it difficult to acquire high-resolution images for supervised learning methods. To address this problem, in this work, we propose STRESS (Spatio-Temporal Resolution Enhancement with Simulated Scans), a self-supervised super-resolution framework for dynamic fetal MRI with interleaved slice acquisitions. Our proposed method simulates an interleaved slice acquisition along the high-resolution axis on the originally acquired data to generate pairs of low- and high-resolution images. Then, it trains a super-resolution network by exploiting both spatial and temporal correlations in the MR time series, which is used to enhance the resolution of the original data. Evaluations on both simulated and in utero data show that our proposed method outperforms other self-supervised super-resolution methods and improves image quality, which is beneficial to other downstream tasks and evaluations.
This work tackles the issue of fairness in the context of generative procedures, such as image super-resolution, which entail different definitions from the standard classification setting. Moreover, while traditional group fairness definitions are typically defined with respect to specified protected groups -- camouflaging the fact that these groupings are artificial and carry historical and political motivations -- we emphasize that there are no ground truth identities. For instance, should South and East Asians be viewed as a single group or separate groups? Should we consider one race as a whole or further split by gender? Choosing which groups are valid and who belongs in them is an impossible dilemma and being "fair" with respect to Asians may require being "unfair" with respect to South Asians. This motivates the introduction of definitions that allow algorithms to be \emph{oblivious} to the relevant groupings. We define several intuitive notions of group fairness and study their incompatibilities and trade-offs. We show that the natural extension of demographic parity is strongly dependent on the grouping, and \emph{impossible} to achieve obliviously. On the other hand, the conceptually new definition we introduce, Conditional Proportional Representation, can be achieved obliviously through Posterior Sampling. Our experiments validate our theoretical results and achieve fair image reconstruction using state-of-the-art generative models.
High-resolution satellite imagery has proven useful for a broad range of tasks, including measurement of global human population, local economic livelihoods, and biodiversity, among many others. Unfortunately, high-resolution imagery is both infrequently collected and expensive to purchase, making it hard to efficiently and effectively scale these downstream tasks over both time and space. We propose a new conditional pixel synthesis model that uses abundant, low-cost, low-resolution imagery to generate accurate high-resolution imagery at locations and times in which it is unavailable. We show that our model attains photo-realistic sample quality and outperforms competing baselines on a key downstream task -- object counting -- particularly in geographic locations where conditions on the ground are changing rapidly.
We propose a manifold matching approach to generative models which includes a distribution generator (or data generator) and a metric generator. In our framework, we view the real data set as some manifold embedded in a high-dimensional Euclidean space. The distribution generator aims at generating samples that follow some distribution condensed around the real data manifold. It is achieved by matching two sets of points using their geometric shape descriptors, such as centroid and $p$-diameter, with learned distance metric; the metric generator utilizes both real data and generated samples to learn a distance metric which is close to some intrinsic geodesic distance on the real data manifold. The produced distance metric is further used for manifold matching. The two networks are learned simultaneously during the training process. We apply the approach on both unsupervised and supervised learning tasks: in unconditional image generation task, the proposed method obtains competitive results compared with existing generative models; in super-resolution task, we incorporate the framework in perception-based models and improve visual qualities by producing samples with more natural textures. Experiments and analysis demonstrate the feasibility and effectiveness of the proposed framework.