Improve Reconstruction Accuracy Research Articles

Convolutional Neural Network-Based Multiple-Rate Compressive Sensing for Massive MIMO CSI Feedback: Design, Simulation, and Analysis

Massive multiple-input multiple-output (MIMO) is a promising technology to increase link capacity and energy efficiency. However, these benefits are based on available channel state information (CSI) at the base station (BS). Therefore, user equipment (UE) needs to keep on feeding CSI back to the BS, thereby consuming precious bandwidth resource. Large-scale antennas at the BS for massive MIMO seriously increase this overhead. In this paper, we propose a multiple-rate compressive sensing neural network framework to compress and quantize the CSI. This framework not only improves reconstruction accuracy but also decreases storage space at the UE, thus enhancing the system feasibility. Specifically, we establish two network design principles for CSI feedback, propose a new network architecture, CsiNet+, according to these principles, and develop a novel quantization framework and training strategy. Next, we further introduce two different variable-rate approaches, namely, SM-CsiNet+ and PM-CsiNet+, which decrease the parameter number at the UE by 38.0% and 46.7%, respectively. Experimental results show that CsiNet+ outperforms the state-of-the-art network by a margin but only slightly increases the parameter number. We also investigate the compression and reconstruction mechanism behind deep learning-based CSI feedback methods via parameter visualization, which provides a guideline for subsequent research.

Simultaneous fast joint sparse recovery for WSN and IoT applications

Joint sparse recovery aims to recover a number of sparse signals having joint sparsity from multiple compressed measurements. Such a problem is finding increasing applications in wireless sensor networks (WSN) and Internet of Things (IoT) where multiple sensors collect measurements. However, existing algorithms lack both reconstruction accuracy and speed at the same time. In this study, the authors propose a joint sparse recovery algorithm, simultaneous fast matching pursuit (SFMP), which exploits some of the concepts developed for the fast matching pursuit (FMP) algorithm. SFMP achieves significant improvement in reconstruction time and speed compared to other related existing algorithms. In contrast to related algorithms, support selection is performed efficiently as the number of selected atoms is adapted from an iteration to another. Furthermore, signal estimation is performed avoiding large matrix inversion as in related algorithms. Moreover, simultaneously pruning the estimated signals, results in removing incorrectly selected ones. Due to the efficient selection strategy and the simultaneous pruning operation, the algorithm shows significant improvement in reconstruction accuracy from noisy measurements. SFMP achieves significant speed improvement over simultaneous orthogonal matching pursuit, and significant accuracy improvement over simultaneous compressive sampling matching pursuit, requiring a much smaller number of measurements.

Fidelity imposed network edit (FINE) for solving ill-posed image reconstruction

Deep learning (DL) is increasingly used to solve ill-posed inverse problems in medical imaging, such as reconstruction from noisy and/or incomplete data, as DL offers advantages over conventional methods that rely on explicit image features and hand engineered priors. However, supervised DL-based methods may achieve poor performance when the test data deviates from the training data, for example, when it has pathologies not encountered in the training data. Furthermore, DL-based image reconstructions do not always incorporate the underlying forward physical model, which may improve performance. Therefore, in this work we introduce a novel approach, called fidelity imposed network edit (FINE), which modifies the weights of a pre-trained reconstruction network for each case in the testing dataset. This is achieved by minimizing an unsupervised fidelity loss function that is based on the forward physical model. FINE is applied to two important inverse problems in neuroimaging: quantitative susceptibility mapping (QSM) and under-sampled image reconstruction in MRI. Our experiments demonstrate that FINE can improve reconstruction accuracy.

Time-domain signal reconstruction of vehicle interior noise based on deep learning and compressed sensing techniques

During vehicle driving, the noise signal of passenger ear-sides is affected by many sound sources. Meanwhile, the composition and generation mechanism of these sound sources are complicated. The application of traditional data-driven technology in the noise signal reconstruction process of passenger ear-sides often requires complex signal processing and prior knowledge, thereby limiting its application in signal reconstruction. Thus, a multivariable-based vehicle interior noise time-domain signal reconstruction (MTSR) algorithm based on compressed sensing and deep learning is proposed to address such limitation. Raw data are compressed to acquire samples using the proposed algorithm to reduce the amount of data and realize the adaptive extraction of signal features. A deep neural network model for the noise signal reconstruction of passenger ear-sides is established on the basis of multisource signals of the compressed domain, which is pretrained by a restricted Boltzmann machine for improved reconstruction accuracy. The recovery compression signal method realizes the time-domain signal reconstruction of the passenger ear-sides. The effectiveness of the proposed MTSR algorithm is validated using noise signal sources collected from a vehicle. Compared with the different reconstruction models, the proposed algorithm is superior in reconstruction accuracy and time consumption.

An End-to-End Approach to Reconstructing 3D Model From Image Set

Large-scale 3D reconstruction from imagery has received much attention from the computer vision community. However, recovering 3D structures from 2D images is a notoriously complex process that requires expertise with often limited results. This paper presents an end-to-end 3D reconstruction system that can produce high-quality 3D models from a set of unordered image collections. Our workflow is a typical 3D reconstruction architecture that consists of structure from motion (SFM), multi-view stereo (MVS), and surface reconstruction, and can automatically recover desirable 3D models without any interactive operations. Finally, a comprehensive experiment is conducted on several benchmark datasets to assess the presented system. Experimental results show that the presented system achieves significant improvements in reconstruction accuracy and completeness over the existing state-of-the-art approach.

Multi-Objective Sparse Reconstruction With Transfer Learning and Localized Regularization

Multi-objective sparse reconstruction methods have shown strong potential in sparse reconstruction. However, most methods are computationally expensive due to the requirement of excessive functional evaluations. Most of these methods adopt arbitrary regularization values for iterative thresholding-based local search, which hardly produces high-precision solutions stably. In this article, we propose a multi-objective sparse reconstruction scheme with novel techniques of transfer learning and localized regularization. Firstly, we design a knowledge transfer operator to reuse the search experience from previously solved homogeneous or heterogeneous sparse reconstruction problems, which can significantly accelerate the convergence and improve the reconstruction quality. Secondly, we develop a localized regularization strategy for iterative thresholding-based local search, which uses systematically designed independent regularization values according to decomposed subproblems. The strategy can lead to improved reconstruction accuracy. Therefore, our proposed scheme is more computationally efficient and accurate, compared to existing multi-objective sparse reconstruction methods. This is validated by extensive experiments on simulated signals and benchmark problems.

Optoacoustic model-based inversion using anisotropic adaptive total-variation regularization

In optoacoustic tomography, image reconstruction is often performed with incomplete or noisy data, leading to reconstruction errors. Significant improvement in reconstruction accuracy may be achieved in such cases by using nonlinear regularization schemes, such as total-variation minimization and L1-based sparsity-preserving schemes. In this paper, we introduce a new framework for optoacoustic image reconstruction based on adaptive anisotropic total-variation regularization, which is more capable of preserving complex boundaries than conventional total-variation regularization. The new scheme is demonstrated in numerical simulations on blood-vessel images as well as on experimental data and is shown to be more capable than the total-variation-L1 scheme in enhancing image contrast.

Spatial-spectral feature based approach towards convolutional sparse coding of hyperspectral images

The sparse coding approaches, which learn over-complete bases for efficiently representing a given set of data, are being extensively studied for various image analysis tasks. However, since sparse analysis of individual bands does not consider inter-spectral characteristics, adaptation of these approaches to hyperspectral domain has been quite limited. In this regard, we investigate convolutional frameworks for simultaneously learning sparse codes and dictionaries so as to efficiently represent the spectral–spatial characteristics of hypercubes. This study also reviews prominent convolutional sparse coding algorithms relevant to hyperspectral datasets. Unlike conventional approaches, in order to address the issues due to huge spectral dimension and high inter-band correlation, the proposed approaches implement sparse coding in a convolutional encoder–decoder framework. Two architectures, one based on shrinkage thresholding algorithm and other based on structured sparse regularization, are proposed. These architectures adopt 3D convolutions in such a way as to model both spectral and spatial features. Although Architecture-1 gives better results than other prominent approaches, its accuracy deteriorates with increase in spectral dimension of input. This can be attributed to the bottleneck due to shrinkage thresholding operation. To address the issue, Architecture-2 adopts structured sparse regularizations over filter weights and feature tensors. Results indicate that the proposed approaches achieve significant improvement in reconstruction accuracy when compared to the existing approaches. The frameworks are also evaluated with respect to super-resolution and dimensionality reduction of hyperspectral datasets. From different experiments, it is observed that the simultaneous usage of 3D and 2D filters along with sparse regularizations significantly improve learning capability of the networks.

Cascaded Deep Networks With Multiple Receptive Fields for Infrared Image Super-Resolution

Infrared images have a wide range of military and civilian applications, including night vision, surveillance, and robotics. However, high-resolution infrared detectors are difficult to fabricate and their manufacturing cost is expensive. In this paper, we present a cascaded architecture of deep neural networks with multiple receptive fields to increase the spatial resolution of infrared images by a large scale factor ( $\times 8$ ). Instead of reconstructing a high-resolution image from its low-resolution version using a single complex deep network, the key idea of our approach is to set up a mid-point (scale $\times 2$ ) between scale $\times 1$ and $\times 8$ such that lost information can be divided into two components. Lost information within each component contains similar patterns thus can be more accurately recovered even using a simpler deep network. In our proposed cascaded architecture, two consecutive deep networks with different receptive fields are jointly trained through a multi-scale loss function. The first network with a large receptive field is applied to recover large-scale structure information, while the second one uses a relatively smaller receptive field to reconstruct small-scale image details. Our proposed method is systematically evaluated using realistic infrared images. Compared with state-of-the-art super-resolution methods, our proposed cascaded approach achieves improved reconstruction accuracy using significantly fewer parameters.

Region-of-interest undersampled MRI reconstruction: A deep convolutional neural network approach

Compressive sensing enables fast magnetic resonance imaging (MRI) reconstruction with undersampled k-space data. However, in most existing MRI reconstruction models, the whole MR image is targeted and reconstructed without taking specific tissue regions into consideration. This may fails to emphasize the reconstruction accuracy on important and region-of-interest (ROI) tissues for diagnosis. In some ROI-based MRI reconstruction models, the ROI mask is extracted by human experts in advance, which is laborious when the MRI datasets are too large. In this paper, we propose a deep neural network architecture for ROI MRI reconstruction called ROIRecNet to improve reconstruction accuracy of the ROI regions in under-sampled MRI. In the model, we obtain the ROI masks by feeding an initially reconstructed MRI from a pre-trained MRI reconstruction network (RecNet) to a pre-trained MRI segmentation network (ROINet). Then we fine-tune the RecNet with a binary weighted ℓ2 loss function using the produced ROI mask. The resulting ROIRecNet can offer more focus on the ROI. We test the model on the MRBrainS13 dataset with different brain tissues being ROIs. The experiment shows the proposed ROIRecNet can significantly improve the reconstruction quality of the region of interest.

Interactive hand pose estimation using a stretch-sensing soft glove

We propose a stretch-sensing soft glove to interactively capture hand poses with high accuracy and without requiring an external optical setup. We demonstrate how our device can be fabricated and calibrated at low cost, using simple tools available in most fabrication labs. To reconstruct the pose from the capacitive sensors embedded in the glove, we propose a deep network architecture that exploits the spatial layout of the sensor itself. The network is trained only once, using an inexpensive off-the-shelf hand pose reconstruction system to gather the training data. The per-user calibration is then performed on-the-fly using only the glove. The glove's capabilities are demonstrated in a series of ablative experiments, exploring different models and calibration methods. Comparing against commercial data gloves, we achieve a 35% improvement in reconstruction accuracy.

Orthogonal tensor dictionary learning for accelerated dynamic MRI.

A direct application of the compressed sensing (CS) theory to dynamic magnetic resonance imaging (MRI) reconstruction needs vectorization or matricization of the dynamic MRI data, which is composed of a stack of 2D images and can be naturally regarded as a tensor. This 1D/2D model may destroy the inherent spatial structure property of the data. An alternative way to exploit the multidimensional structure in dynamic MRI is to employ tensor decomposition for dictionary learning, that is, learning multiple dictionaries along each dimension (mode) and sparsely representing the multidimensional data with respect to the Kronecker product of these dictionaries. In this work, we introduce a novel tensor dictionary learning method under an orthonormal constraint on the elementary matrix of the tensor dictionary for dynamic MRI reconstruction. The proposed algorithm alternates sparse coding, tensor dictionary learning, and updating reconstruction, and each corresponding subproblem is efficiently solved by a closed-form solution. Numerical experiments on phantom and synthetic data show significant improvements in reconstruction accuracy and computational efficiency obtained by the proposed scheme over the existing method that uses the 1D/2D model with overcomplete dictionary learning. Graphical abstract Fig. 1 Comparison between (a) the traditional method and (b) the proposed method based on dictionary learning for dynamic MRI reconstruction.

Unsupervised binocular depth prediction network for laparoscopic surgery.

Minimally invasive surgery (MIS) is characterized by less trauma, shorter recovery time, and lower postoperative infection rate. The two-dimensional (2D) laparoscopic imaging lacks depth perception and does not provide quantitative depth information, thereby limiting precise and complex surgical operations. Three-dimensional (3D) laparoscopic imaging provides surgeons depth perception. This study aims to 3D reconstruction of the surgical scene based on the disparity map generated by the depth estimation algorithm. An unsupervised learning autoencoder method was proposed to calculate the accurate disparity with a 101-layer residual convolutional network. The loss function included three parts: left-right consistency loss, structure similarity loss, and reconstruction error loss, the combination can improve reconstruction accuracy and robustness. The method was validated on a Hamlyn Center Laparoscopic/Endoscopic Video Dataset. The structural similarity index (SSIM) is 0.8349 ± 0.0523 and the peak signal-to-noise ratio (PSNR) is 14.4957 ± 1.9676. The depth prediction network has high accuracy and robustness. The average time to produce each disparity map is about 16 ms. The experimental result shows that the proposed depth estimation method can offer dense disparity map, and can meet surgical real-time requirement. Future work will focus on network structure optimization and loss function design, transfer learning to improve the robustness and accuracy further.

Improving reconstruction of tunnel lining defects from ground-penetrating radar profiles by multi-scale inversion and bi-parametric full-waveform inversion

Complex irregular defects of tunnel linings under complicated geological conditions cannot be accurately reconstructed with the traditional full-waveform inversion (FWI) method due to their irregular geometrical characteristics and complex dielectric properties. Because of extensive inversion calculations and high memory requirements, the traditional FWI method is very sensitive to the initial model and plunge into a local minimum or cycle skipping. To solve this problem, a novel ground-penetrating radar (GPR) FWI method involving two parameters (i.e. permittivity and conductivity) is proposed for improving reconstruction accuracy of lining defects using the total variation (TV) regularization. First, the Delaunay unstructured triangular mesh in finite element time-domain (FETD) method is employed to perform GPR forward modeling, and then the total-variation model constraint and multi-scale inversion strategy are implemented during execution of the conjugate gradient (CG) algorithm, which facilitates the quick search for the global optimal minimum value, thus guaranteeing the avoidance of the ill-posed problem during the inversion process. Accordingly, the detailed features of lining defects can be characterized and reconstructed even for those complicated geological conditions, more specifically, the results show that, with fewer iterations (up to 59 times less), the proposed method present a lower reconstruction error both on permittivity (up to 12.05% lower) and conductivity (up to 7.35% lower). From a comparison of the inversion results and the model, it can be concluded that the proposed FWI algorithm can effectively eliminate the non-physical oscillation and artifacts in the image reconstruction, which may significantly improving the accuracy of defects interpretation and assessing the severity of complex defects.

Compressing unstructured mesh data from simulations using machine learning

The amount of data output from a computer simulation has grown to terabytes and petabytes as increasingly complex simulations are being run on massively parallel systems. As we approach exaflop computing in the next decade, it is expected that the I/O subsystem will not be able to write out these large volumes of data. In this paper, we explore the use of machine learning to compress the data before it is written out. Despite the computational constraints that limit us to using very simple learning algorithms, our results show that machine learning is a viable option for compressing unstructured data. We demonstrate that by simply using a better sampling algorithm to generate the training set, we can obtain more accurate results compared to random sampling, but at no extra cost. Further, by carefully selecting and incorporating points with high prediction error, we can improve reconstruction accuracy without sacrificing the compression rate.

An Adaptive Sampling Scheme via Approximate Volume Sampling for Fingerprint-Based Indoor Localization

In recent years Wi-Fi fingerprinting has attracted much attention in indoor localization because of the availability of high-quality signal and pervasive deployment of wireless LANs. For fingerprint-based localization, however, offline site survey is usually time-consuming and labor-intensive. Therefore, reducing the burden of offline site survey becomes an important issue for fingerprint-based indoor localization. In this paper, using a low-tubal-rank tensor to model Wi-Fi fingerprints of all reference points (RPs), we propose an adaptive sampling scheme via approximate volume sampling to improve reconstruction accuracy of radio map with reduced expenditure. We propose a rank-increasing strategy to effectively estimate the rank of the underlying fingerprint tensor to alleviate the computation burden for tensor completion. We provide a theoretical foundation to analyze the proposed scheme and derive the performance bounds in terms of sample complexity and reconstruction error. We prove that the proposed scheme can achieve a relative error guarantee. Finally, we validate the effectiveness of the proposed scheme through extensive simulations using both synthetic and real datasets. The simulation results demonstrate that the proposed scheme is able to not only reduce reconstruction error and improve localization accuracy but also reduce running time compared to the state-of-the-art schemes.

Multi-Resolution Graph Based Volumetric Cortical Basis Functions From Local Anatomic Features.

Modern clinical MRI collects millimeter scale anatomic information, but scalp electroencephalography source localization is ill posed, and cannot resolve individual sources at that resolution. Dimensionality reduction in the space of cortical sources is needed to improve computational and storage complexity, yet volumetric methods still employ simplistic grid coarsening that eliminates fine scale anatomic structure. We present an approach to extend near-arbitrary spatial scaling to volumetric localization. Starting from a voxelwise brain parcellation, sub-parcels are identified from local cortical connectivity with an iterated graph cut approach. Spatial basis functions in each parcel are constructed using either a decomposition of the local leadfield matrix or spectral basis functions of local cortical connectivity graphs. We present quantitative evaluation with extensive simulations and use multiple sets of real data to highlight how parameter changes impact computed reconstructions. Our results show that volumetric basis functions can improve accuracy by as much as 30%, while reducing computational complexity by over two orders of magnitude. In real data from epilepsy surgical candidates, accurate localization of seizure onset regions is demonstrated. Spatial dimensionality reduction with volumetric basis functions improves reconstruction accuracy while reducing computational complexity. Near-arbitrary spatial dimensionality reduction will enable volumetric reconstruction with modern computationally intensive algorithms and anatomically driven multi-resolution methods.

Strain Gauges Based 3D Shape Monitoring of Beam Structures Using Finite Width Gauge Model

This paper presents a new approach validated experimentally to reconstruct with strain gauges the deformed shape of a straight beam with circular cross section. It is based on a novel beam-specific strain gauge model that improves the strain measurement by taking into account the width of the gauges. These improved strain measurements are used by a 3D finite strain large displacement beam shape reconstruction method to recover the deformed shape iteratively. The whole reconstruction approach has been validated experimentally with 3D deformations of a beam instru-mented with strain gauges. Results show that the strain gauge model developed improves reconstruction accuracy and that beam reconstruction can be achieved effectively .

Real-Time Dense Monocular SLAM With Online Adapted Depth Prediction Network

Considerable advances have been achieved in estimating the depth map from a single image via convolutional neural networks (CNNs) during the past few years. Combining depth prediction from CNNs with conventional monocular simultaneous localization and mapping (SLAM) is promising for accurate and dense monocular reconstruction, in particular addressing the two long-standing challenges in conventional monocular SLAM: low map completeness and scale ambiguity. However, depth estimated by pretrained CNNs usually fails to achieve sufficient accuracy for environments of different types from the training data, which are common for certain applications such as obstacle avoidance of drones in unknown scenes. Additionally, inaccurate depth prediction of CNN could yield large tracking errors in monocular SLAM. In this paper, we present a real-time dense monocular SLAM system, which effectively fuses direct monocular SLAM with an online-adapted depth prediction network for achieving accurate depth prediction of scenes of different types from the training data and providing absolute scale information for tracking and mapping. Specifically, on one hand, tracking pose (i.e., translation and rotation) from direct SLAM is used for selecting a small set of highly effective and reliable training images, which acts as ground truth for tuning the depth prediction network on-the-fly toward better generalization ability for scenes of different types. A stage-wise Stochastic Gradient Descent algorithm with a selective update strategy is introduced for efficient convergence of the tuning process. On the other hand, the dense map produced by the adapted network is applied to address scale ambiguity of direct monocular SLAM which in turn improves the accuracy of both tracking and overall reconstruction. The system with assistance of both CPUs and GPUs, can achieve real-time performance with progressively improved reconstruction accuracy. Experimental results on public datasets and live application to obstacle avoidance of drones demonstrate that our method outperforms the state-of-the-art methods with greater map completeness and accuracy, and a smaller tracking error.

Fast and Accurate Super-Resolution of FY-2 Infrared Cloud Images via Multi-Scale Fusion Network

This paper proposes an effective method to improve the spatial resolution of FengYun-2 (FY-2) infrared cloud images via deep convolutional neural networks. The proposed model consists of four parts: shallow feature representation block, stacked multi-scale fusion blocks, global feature fusion block, and feature reconstruction block. The multi-scale fusion block combines dilated convolution, local feature fusion and local residual learning to extract multi-scale local features from the original low-resolution image directly. Then these local features are all merged by the global feature fusion block to reconstruct the residual representations in high-resolution space. For training and testing, we have specially built a dataset of infrared cloud images. We evaluated the proposed method both on simulated and real data. Experimental results demonstrate that the proposed approach achieves improved reconstruction accuracy than the state-of-the-art methods. Besides, the concise structure of the proposed model enables it to be applicable in practice.