Computed tomography is widely used to examine internal structures in a non-destructive manner. To obtain high-quality reconstructions, one typically has to acquire a densely sampled trajectory to avoid angular undersampling. However, many scenarios require a sparse-view measurement leading to streak-artifacts if unaccounted for. Current methods do not make full use of the domain-specific information, and hence fail to provide reliable reconstructions for highly undersampled data. We present a novel framework for sparse-view tomography by decoupling the reconstruction into two steps: First, we overcome its ill-posedness using a super-resolution network, SIN, trained on the sparse projections. The intermediate result allows for a closed-form tomographic reconstruction with preserved details and highly reduced streak-artifacts. Second, a refinement network, PRN, trained on the reconstructions reduces any remaining artifacts. We further propose a light-weight variant of the perceptual-loss that enhances domain-specific information, boosting restoration accuracy. Our experiments demonstrate an improvement over current solutions by 4 dB.
Generative Adversarial Networks (GANs) advance face synthesis through learning the underlying distribution of observed data. Despite the high-quality generated faces, some minority groups can be rarely generated from the trained models due to a biased image generation process. To study the issue, we first conduct an empirical study on a pre-trained face synthesis model. We observe that after training the GAN model not only carries the biases in the training data but also amplifies them to some degree in the image generation process. To further improve the fairness of image generation, we propose an interpretable baseline method to balance the output facial attributes without retraining. The proposed method shifts the interpretable semantic distribution in the latent space for a more balanced image generation while preserving the sample diversity. Besides producing more balanced data regarding a particular attribute (e.g., race, gender, etc.), our method is generalizable to handle more than one attribute at a time and synthesize samples of fine-grained subgroups. We further show the positive applicability of the balanced data sampled from GANs to quantify the biases in other face recognition systems, like commercial face attribute classifiers and face super-resolution algorithms.
Convolutional neural networks are the most successful models in single image super-resolution. Deeper networks, residual connections, and attention mechanisms have further improved their performance. However, these strategies often improve the reconstruction performance at the expense of considerably increasing the computational cost. This paper introduces a new lightweight super-resolution model based on an efficient method for residual feature and attention aggregation. In order to make an efficient use of the residual features, these are hierarchically aggregated into feature banks for posterior usage at the network output. In parallel, a lightweight hierarchical attention mechanism extracts the most relevant features from the network into attention banks for improving the final output and preventing the information loss through the successive operations inside the network. Therefore, the processing is split into two independent paths of computation that can be simultaneously carried out, resulting in a highly efficient and effective model for reconstructing fine details on high-resolution images from their low-resolution counterparts. Our proposed architecture surpasses state-of-the-art performance in several datasets, while maintaining relatively low computation and memory footprint.
Machine learning models are commonly trained end-to-end and in a supervised setting, using paired (input, output) data. Examples include recent super-resolution methods that train on pairs of (low-resolution, high-resolution) images. However, these end-to-end approaches require re-training every time there is a distribution shift in the inputs (e.g., night images vs daylight) or relevant latent variables (e.g., camera blur or hand motion). In this work, we leverage state-of-the-art (SOTA) generative models (here StyleGAN2) for building powerful image priors, which enable application of Bayes' theorem for many downstream reconstruction tasks. Our method, Bayesian Reconstruction through Generative Models (BRGM), uses a single pre-trained generator model to solve different image restoration tasks, i.e., super-resolution and in-painting, by combining it with different forward corruption models. We keep the weights of the generator model fixed, and reconstruct the image by estimating the Bayesian maximum a-posteriori (MAP) estimate over the input latent vector that generated the reconstructed image. We further use variational inference to approximate the posterior distribution over the latent vectors, from which we sample multiple solutions. We demonstrate BRGM on three large and diverse datasets: (i) 60,000 images from the Flick Faces High Quality dataset (ii) 240,000 chest X-rays from MIMIC III and (iii) a combined collection of 5 brain MRI datasets with 7,329 scans. Across all three datasets and without any dataset-specific hyperparameter tuning, our simple approach yields performance competitive with current task-specific state-of-the-art methods on super-resolution and in-painting, while being more generalisable and without requiring any training. Our source code and pre-trained models are available online: https://razvanmarinescu.github.io/brgm/.
Advances in face rotation, along with other face-based generative tasks, are more frequent as we advance further in topics of deep learning. Even as impressive milestones are achieved in synthesizing faces, the importance of preserving identity is needed in practice and should not be overlooked. Also, the difficulty should not be more for data with obscured faces, heavier poses, and lower quality. Existing methods tend to focus on samples with variation in pose, but with the assumption data is high in quality. We propose a generative adversarial network (GAN) -based model to generate high-quality, identity preserving frontal faces from one or multiple low-resolution (LR) faces with extreme poses. Specifically, we propose SuperFront-GAN (SF-GAN) to synthesize a high-resolution (HR), frontal face from one-to-many LR faces with various poses and with the identity-preserved. We integrate a super-resolution (SR) side-view module into SF-GAN to preserve identity information and fine details of the side-views in HR space, which helps model reconstruct high-frequency information of faces (i.e., periocular, nose, and mouth regions). Moreover, SF-GAN accepts multiple LR faces as input, and improves each added sample. We squeeze additional gain in performance with an orthogonal constraint in the generator to penalize redundant latent representations and, hence, diversify the learned features space. Quantitative and qualitative results demonstrate the superiority of SF-GAN over others.
A combination of cloud-based deep learning (DL) algorithms with portable/wearable (P/W) devices has been developed as a smart heath care system to support automatic cardiac arrhythmias (CAs) classification using electrocardiography (ECG). However, long-term and continuous ECG monitoring is challenging because of limitations of batteries and transmission bandwidth of P/W devices while incorporated with consumer electronics (CE). A feasible approach to address this challenge is to decrease sampling rates. However, low sampling rates lead to low-resolution signals that hinder the CAs classification performance. In this study, we propose a DL-based ECG signal super-resolution framework (called SRECG) to enhance low-resolution ECG signals by jointly considering the accuracies when applied to the DL-based high-resolution multiclass classifier (HMC) of CAs. In our experiments, we downsampled the ECG signals from the CPSC2018 dataset and evaluated their HMC accuracies with and without the SRECG. Experimental results show that SRECG can well improve the HMC accuracies as compared to traditional interpolation methods. Moreover, approximately half of the CAs classification accuracies of HMC were maintained within the enhanced ECG signals by SRECG. The promising results confirm that SRECG can be suitably used to enhance low-resolution ECG signals from P/W devices with CE to improve their cloud-based HMC performances.
Block-sparse regularization is already well-known in active thermal imaging and is used for multiple measurement based inverse problems. The main bottleneck of this method is the choice of regularization parameters which differs for each experiment. To avoid time-consuming manually selected regularization parameter, we propose a learned block-sparse optimization approach using an iterative algorithm unfolded into a deep neural network. More precisely, we show the benefits of using a learned block iterative shrinkage thresholding algorithm that is able to learn the choice of regularization parameters. In addition, this algorithm enables the determination of a suitable weight matrix to solve the underlying inverse problem. Therefore, in this paper we present the algorithm and compare it with state of the art block iterative shrinkage thresholding using synthetically generated test data and experimental test data from active thermography for defect reconstruction. Our results show that the use of the learned block-sparse optimization approach provides smaller normalized mean square errors for a small fixed number of iterations than without learning. Thus, this new approach allows to improve the convergence speed and only needs a few iterations to generate accurate defect reconstruction in photothermal super resolution imaging.
Image Super-Resolution (SR) provides a promising technique to enhance the image quality of low-resolution optical sensors, facilitating better-performing target detection and autonomous navigation in a wide range of robotics applications. It is noted that the state-of-the-art SR methods are typically trained and tested using single-channel inputs, neglecting the fact that the cost of capturing high-resolution images in different spectral domains varies significantly. In this paper, we attempt to leverage complementary information from a low-cost channel (visible/depth) to boost image quality of an expensive channel (thermal) using fewer parameters. To this end, we first present an effective method to virtually generate pixel-wise aligned visible and thermal images based on real-time 3D reconstruction of multi-modal data captured at various viewpoints. Then, we design a feature-level multispectral fusion residual network model to perform high-accuracy SR of thermal images by adaptively integrating co-occurrence features presented in multispectral images. Experimental results demonstrate that this new approach can effectively alleviate the ill-posed inverse problem of image SR by taking into account complementary information from an additional low-cost channel, significantly outperforming state-of-the-art SR approaches in terms of both accuracy and efficiency.
Video transmission applications (e.g., conferencing) are gaining momentum, especially in times of global health pandemic. Video signals are transmitted over lossy channels, resulting in low-quality received signals. To restore videos on recipient edge devices in real-time, we introduce an efficient video restoration network, EVRNet. EVRNet efficiently allocates parameters inside the network using alignment, differential, and fusion modules. With extensive experiments on video restoration tasks (deblocking, denoising, and super-resolution), we demonstrate that EVRNet delivers competitive performance to existing methods with significantly fewer parameters and MACs. For example, EVRNet has 260 times fewer parameters and 958 times fewer MACs than enhanced deformable convolution-based video restoration network (EDVR) for 4 times video super-resolution while its SSIM score is 0.018 less than EDVR. We also evaluated the performance of EVRNet under multiple distortions on unseen dataset to demonstrate its ability in modeling variable-length sequences under both camera and object motion.
Despite the great success of deep model on Hyperspectral imagery (HSI) super-resolution(SR) for simulated data, most of them function unsatisfactory when applied to the real data, especially for unsupervised HSI SR methods. One of the main reason comes from the fact that the predefined degeneration models (e.g. blur in spatial domain) utilized by most HSI SR methods often exist great discrepancy with the real one, which results in these deep models overfit and ultimately degrade their performance on real data. To well mitigate such a problem, we explore the unsupervised blind HSI SR method. Specifically, we investigate how to effectively obtain the degeneration models in spatial and spectral domain, respectively, and makes them can well compatible with the fusion based SR reconstruction model. To this end, we first propose an alternating optimization based deep framework to estimate the degeneration models and reconstruct the latent image, with which the degeneration models estimation and HSI reconstruction can mutually promotes each other. Then, a meta-learning based mechanism is further proposed to pre-train the network, which can effectively improve the speed and generalization ability adapting to different complex degeneration. Experiments on three benchmark HSI SR datasets report an excellent superiority of the proposed method on handling blind HSI fusion problem over other competing methods.
Video super-resolution (VSR) approaches tend to have more components than the image counterparts as they need to exploit the additional temporal dimension. Complex designs are not uncommon. In this study, we wish to untangle the knots and reconsider some most essential components for VSR guided by four basic functionalities, i.e., Propagation, Alignment, Aggregation, and Upsampling. By reusing some existing components added with minimal redesigns, we show a succinct pipeline, BasicVSR, that achieves appealing improvements in terms of speed and restoration quality in comparison to many state-of-the-art algorithms. We conduct systematic analysis to explain how such gain can be obtained and discuss the pitfalls. We further show the extensibility of BasicVSR by presenting an information-refill mechanism and a coupled propagation scheme to facilitate information aggregation. The BasicVSR and its extension, IconVSR, can serve as strong baselines for future VSR approaches.
Reconstruction of undersampled periodic signals of unknown period is an important signal processing operation. It is especially difficult operation when the sequences of samples are short and no information on the inter-sequence time distances is given. For such a case, there exist some algorithms that allow for approximation of the sampled signal. However, these algorithms require either bandlimitedness of the signal, or noiseless samples. In this paper, we propose a novel algorithm which does not require the signal to be bandlimited and it can cope with additive noise in the samples. The algorithm is illustrated and validated with real data.
Complex ADMM-Net, a complex-valued neural network architecture inspired by the alternating direction method of multipliers (ADMM), is designed for interference removal in super-resolution stepped frequency radar angle-range-doppler imaging. Tailored to an uncooperative scenario wherein a MIMO radar shares spectrum with communications, the ADMM-Net recovers the radar image---which is assumed to be sparse---and simultaneously removes the communication interference, which is modeled as sparse in the frequency domain owing to spectrum underutilization. The scenario motivates an $\ell_1$-minimization problem whose ADMM iteration, in turn, undergirds the neural network design, yielding a set of generalized ADMM iterations that have learnable hyperparameters and operations. To train the network we use random data generated according to the radar and communication signal models. In numerical experiments ADMM-Net exhibits markedly lower error and computational cost than ADMM and CVX.
Generation of super-resolution (SR) ultrasound (US) images, created from the successive local-ization of individual microbubbles in the circulation, has enabled the visualization of microvascular structure and flow at a level of detail that was not possible previously. Despite rapid progress, tradeoffs between spatial and temporal resolution may challenge the translation of this promising technology to the clinic. To temper these trade-offs, we propose a method based on morphological image reconstriction. This method can extract from ultrafast contrast-enhanced ultrasound (CEUS) images hundreds of microbubble peaks per image (312-by-180 pixels) with intensity values varying by an order of magnitude. Specifically, it offers a fourfold increase in the number of peaks detected per frame, requires on the order of 100 ms for processing, and is robust to additive electronic noise (down to 3.6 dB CNR in CEUS images). By integrating this method to a SR framework we demonstrate a 6-fold improvement in spatial resolution, as compared to CEUS, in imaging chicken embryo microvessels. This method that is computationally efficient and, thus, scalable to large data sets, may augment the abilities of SR-US in imaging microvascular structure and function.
Segmentation of the developing fetal brain is an important step in quantitative analyses. However, manual segmentation is a very time-consuming task which is prone to error and must be completed by highly specialized indi-viduals. Super-resolution reconstruction of fetal MRI has become standard for processing such data as it improves image quality and resolution. However, dif-ferent pipelines result in slightly different outputs, further complicating the gen-eralization of segmentation methods aiming to segment super-resolution data. Therefore, we propose using transfer learning with noisy multi-class labels to automatically segment high resolution fetal brain MRIs using a single set of seg-mentations created with one reconstruction method and tested for generalizability across other reconstruction methods. Our results show that the network can auto-matically segment fetal brain reconstructions into 7 different tissue types, regard-less of reconstruction method used. Transfer learning offers some advantages when compared to training without pre-initialized weights, but the network trained on clean labels had more accurate segmentations overall. No additional manual segmentations were required. Therefore, the proposed network has the potential to eliminate the need for manual segmentations needed in quantitative analyses of the fetal brain independent of reconstruction method used, offering an unbiased way to quantify normal and pathological neurodevelopment.
In this paper, we model the signal propagation effect in ultrasonic imaging using Huygens principle and use this model to develop sparse signal representation based imaging techniques for a two-layer object immersed in water. Relying on the fact that the image of interest is sparse, we cast such an array based imaging problem as a sparse signal recovery problem and develop two types of imaging methods, one method uses only one transducer to illuminate the region of interest {and for this case the system is modeled as a single input multiple output (SIMO) system. The second method relies on all transducers to transmit ultrasonic waves into the material under test and in this case the system is modeled as a multiple input multiple output (MIMO) system}. We further extend our work to a scenario where the propagation velocity of the wave in the object under test is not known precisely. {We discuss different techniques such as greedy based algorithms as well as $\ell_1$-norm minimization based approach to solve the proposed sparse signal representation based method. We give an assessment of the computational complexity of the $\ell_1$-norm minimization based approach for the SIMO and the MIMO cases. We further point out the superiority of the $\ell_1$-norm minimization based approach over the greedy based algorithms. Then we give a comprehensive analysis of error for both the greedy based approaches as well as the $\ell_1$-norm minimization based technique for both the SIMO and the MIMO cases. The analysis utilizes tools from two powerful branches of modern analysis, \emph{local analysis in Banach spaces} and \emph{concentration of measure}}. We finally apply our methods to experimental data gathered from a solid test sample immersed in water and show that sparse signal recovery based techniques outperform the conventional methods available in the literature.
Relaxometry studies in preterm and at-term newborns have provided insight into brain microstructure, thus opening new avenues for studying normal brain development and supporting diagnosis in equivocal neurological situations. However, such quantitative techniques require long acquisition times and therefore cannot be straightforwardly translated to in utero brain developmental studies. In clinical fetal brain magnetic resonance imaging routine, 2D low-resolution T2-weighted fast spin echo sequences are used to minimize the effects of unpredictable fetal motion during acquisition. As super-resolution techniques make it possible to reconstruct a 3D high-resolution volume of the fetal brain from clinical low-resolution images, their combination with quantitative acquisition schemes could provide fast and accurate T2 measurements. In this context, the present work demonstrates the feasibility of using super-resolution reconstruction from conventional T2-weighted fast spin echo sequences for 3D isotropic T2 mapping. A quantitative magnetic resonance phantom was imaged using a clinical T2-weighted fast spin echo sequence at variable echo time to allow for super-resolution reconstruction at every echo time and subsequent T2 mapping of samples whose relaxometric properties are close to those of fetal brain tissue. We demonstrate that this approach is highly repeatable, accurate and robust when using six echo times (total acquisition time under 9 minutes) as compared to gold-standard single-echo spin echo sequences (several hours for one single 2D slice).
Since the advent of generative adversarial networks (GANs), various loss functions have been developed and combined to constitute the overall training objective function, in order to improve model performance or for specific learning tasks. For instance, in image enhancement or restoration, there are often several criteria to consider such as signal-noise ratio, smoothness, structures and details. However, when the optimization goal has more than one adversarial loss, balancing multiple losses in the overall function becomes a challenging, critical and time-consuming problem. In this paper, we propose to tackle the problem by means of efficient multi-objective optimization. The proposed HypervolGAN adopts an adapted version of hypervolume maximization method to effectively define the multi-objective training function for GAN. We tested our proposed method on solving single image super-resolution problem. Experiments show that the proposed HypervolGAN is efficient in saving computational time and efforts for fine-tuning weights of various losses, and can generate enhanced samples that have better quality than results given by baseline GANs. The work explores the integration of adversarial learning and optimization techniques, which can benefit not only image processing but also a wide range of applications.
This work revisits coupled tensor decomposition (CTD)-based hyperspectral super-resolution (HSR). HSR aims at fusing a pair of hyperspectral and multispectral images to recover a super-resolution image (SRI). The vast majority of the HSR approaches take a low-rank matrix recovery perspective. The challenge is that theoretical guarantees for recovering the SRI using low-rank matrix models are either elusive or derived under stringent conditions. A couple of recent CTD-based methods ensure recoverability for the SRI under relatively mild conditions, leveraging on algebraic properties of the canonical polyadic decomposition (CPD) and the Tucker decomposition models, respectively. However, the latent factors of both the CPD and Tucker models have no physical interpretations in the context of spectral image analysis, which makes incorporating prior information challenging---but using priors is often essential for enhancing performance in noisy environments. This work employs an idea that models spectral images as tensors following the block-term decomposition model with multilinear rank-$(L_r, L_r, 1)$ terms (i.e., the LL1 model) and formulates the HSR problem as a coupled LL1 tensor decomposition problem. Similar to the existing CTD approaches, recoverability of the SRI is shown under mild conditions. More importantly, the latent factors of the LL1 model can be interpreted as the key constituents of spectral images, i.e., the endmembers' spectral signatures and abundance maps. This connection allows us to easily incorporate prior information for performance enhancement. A flexible algorithmic framework that can work with a series of structural information is proposed to take advantage of the model interpretability. The effectiveness is showcased using simulated and real data.
In radar accurate localization of unmanned aerial vehicle (UAV) swarms, the high density, similar motion parameters, small radar cross-section (RCS), strong noise and far range put forward high requirements on radar resolution and transmitting power. In this paper, by using advantages of the long-time integration (LTI) technique and gridless sparse method, we construct a super-resolution framework for radar accurate localization of UAV swarms without changing radar hardware and system parameters. Thereafter, based on this framework, a range super-resolution method is proposed to realize the radar accurate localization of UAV swarms. Mathematical analyses and numerical simulations are performed and demonstrate that, compared to the keystone transform (KT)-based LTI method, MUSIC-based method and reweighted atomic-norm minimization (RAM)-based method, the range super-resolution method is more robust and practical for radar accurate localization of UAV swarms under the noisy environment. Additionally, the real experiment with X-band radar is also conducted to verify the effectiveness of the range super-resolution method.