1 Introduction
A variety of imaging inverse problems can be discretized to a linear system where is the measured data, the imaging or forward operator, the object being probed by applying (often called the model), and the noise. Depending on the application, the set of plausible reconstructions could model natural, seismic, or biomedical images. In many cases the resulting inverse problem is illposed, either because of the poor conditioning of (a consequence of the underlying physics) or because .
A classical approach to solve illposed inverse problems is to minimize an objective functional regularized via a certain norm (e.g. , , total variation (TV) seminorm) of the model. These methods promote general properties such as sparsity or smoothness of reconstructions, sometimes in combination with learned synthesis or analysis operators, or dictionaries (Sprechmann et al. [2013]).
In this paper, we address situations with very sparse measurement data () so that even a coarse reconstruction of the unknown model is hard to get with traditional regularization schemes. Unlike artifactremoval scenarios where applying a regularized pseudoinverse of the imaging operator already brings out considerable structure, we look at applications where standard techniques cannot produce a reasonable image (Figure 1). This highly unresolved regime is common in geophysics and requires alternative, more involved strategies (Galetti et al. [2017]).
An appealing alternative to classical regularizers is to use deep neural networks. For example, generative models (GANs) based on neural networks have recently achieved impressive results in regularization of inverse problems (
Bora et al. [2018], Lunz et al. [2018]). However, a difficulty in geophysical applications is that there are very few examples of ground truth models available for training (sometimes none at all). Since GANs require many, they cannot be applied to such problems. This suggests to look for methods that are not very sensitive to the training dataset. Conversely, it means that the sought reconstructions are less detailed than what is expected in datarich settings; for an example, see the reconstructions of the Tibetan plateau (Yao et al. [2006]).In this paper, we propose a twostage method to solve illposed inverse problems using random lowdimensional projections and convolutional neural networks. We first decompose the inverse problem into a collection of simpler learning problems of estimating projections into random (but structured) lowdimensional subspaces of piecewiseconstant images. Each projection is easier to learn in terms of generalization error (
Cooper [1995]) thanks to its lower Lipschitz constant.In the second stage, we solve a new linear inverse problem that combines the estimates from the different subspaces. We show that this converts the original problem with possibly nonlocal (often tomographic) measurements into an inverse problem with localized measurements, and that in fact, in expectation over random subspaces the problem becomes a deconvolution. Intuitively, projecting into piecewiseconstant subspaces is equivalent to estimating local averages—a simpler problem than estimating individual pixel values. Combining the local estimates lets us recover the underlying structure. We believe that this technique is of independent interest in addressing inverse problems.
We test our method on linearized seismic traveltime tomography (Bording et al. [1987], Hole [1992]) with sparse measurements and show that it outperforms learned direct inversion in quality of achieved reconstructions, robustness to measurement errors, and (in)sensitivity to the training data. The latter is essential in domains with insufficient ground truth images.
2 Related work
Although neural networks have long been used to address inverse problems (Ogawa et al. [1998], Hoole [1993], Schiller and Doerffer [2010]
), the past few years have seen the number of related deep learning papers grow exponentially. The majority address biomedical imaging (
Güler and Übeylı [2005], Hudson and Cohen [2000]) with several special issues^{1}^{1}1IEEE Transactions on Medical Imaging, May 2016 (Greenspan et al. [2016]); IEEE Signal Processing Magazine, November 2017, January 2018 (Porikli et al. [2017, 2018]). and review papers (Lucas et al. [2018], McCann et al. [2017]) dedicated to the topic. All these papers address reconstruction from subsampled or lowquality data, often motivated by reduced scanning time or lower radiation doses. Beyond biomedical imaging, machine learning techniques are emerging in geophysical imaging (
ArayaPolo et al. [2017], Lewis et al. [2017], Bianco and Gertoft [2017]), though at a slower pace, perhaps partly due to the lack of standard open datasets.Existing methods can be grouped into noniterative methods that learn a feedforward mapping from the measured data (or some standard manipulation such as adjoint or a pseudoinverse) to the model (Jin et al. [2016], Pelt and Batenburg [2013], Zhu et al. [2018], Wang [2016], Antholzer et al. [2017], Han et al. [2016], Zhang et al. [2016]); and iterative energy minimization methods, with either the regularizer being a neural network (Li et al. [2018]), or neural networks replacing various iteration components such as gradients, projectors, or proximal mappings (Kelly et al. [2017], Adler and Öktem [2017b, a], Rick Chang et al. [2017]). These are further related to the notion of plugandplay regularization (Venkatakrishnan et al. [2013]), as well as early uses of neural nets to unroll and adapt standard sparse reconstruction algorithms (Gregor and LeCun [2010], Xin et al. [2016]). An advantage of the first group of methods is that they are fast; an advantage of the second group is that they are better at enforcing data consistency.
Generative models
A rather different take was proposed in the context of compressed sensing where the reconstruction is constrained to lie in the range of a pretrained generative network (Bora et al. [2017, 2018]). Their scheme achieves impressive results on random sensing operators and comes with theoretical guarantees. However, training generative networks requires many examples of ground truth and the method is inherently subject to dataset bias. Here, we focus on a setting where groundtruth samples are very few or impossible to obtain.
There are connections between our work and sketching (Gribonval et al. [2017], Pilanci and Wainwright [2016]) where the learning problem is also simplified by random lowdimensional projections of some object—either the data or the unknown reconstruction itself (Yurtsever et al. [2017]). This also exposes natural connections with learning via random features (Rahimi and Recht [2008, 2009]).
3 Regularization by random mesh projections
The two stages of our method are (i) decomposing a “hard” learning task of directly learning an unstable operator into an ensemble of “easy” tasks of estimating projections of the unknown model into lowdimensional subspaces; and (ii) combining these projection estimates to solve a reformulated inverse problem for . The two stages are summarized in Figure 2. While our method is applicable to continuous and nonlinear settings, we focus on linear finitedimensional inverse problems.
3.1 Decomposing the learning problem
Statistical learning theory tells us that the number of samples required to learn a variate Lipschitz function to a given supnorm accuracy is (Cooper [1995]). While this result is proved for scalarvalued multivariate maps, it is reasonable to expect the same scaling in
to hold for vectorvalued maps. This motivates us to study Lipschitz properties of the projected inverse maps.
We wish to reconstruct , an pixel image from where is large (we think of as an discrete image). We assume that the map from to is injective so that it is invertible on its range, and that there exists an Lipschitz (generally nonlinear) inverse ,
In order for the injectivity assumption to be reasonable, we assume that is a lowdimensional manifold embedded in of dimension at most , where is the number of measurements. Since we are in finite dimension, injectivity implies the existence of (Stefanov and Uhlmann [2009]). Due to illposedness, is typically large.
Consider now the map from the data to a projection of the model into some dimensional subspace , where . Note that this map exists by construction (since is injective on ), and that it must be nonlinear. To see this, note that the only consistent^{2}^{2}2Consistent meaning that if already lives in , then the map should return . linear map acting on is an oblique, rather than an orthogonal projection on (cf. Section 2.4 in Vetterli et al. [2014]). We explain this in more detail in Appendix A.
Denote the projection by and assume is chosen uniformly at random.^{3}^{3}3One way to construct the corresponding projection matrix is as , where is a matrix with standard iid Gaussian entries. We want to evaluate the expected Lipschitz constant of the map from to , noting that it can be written as :
where the first inequality is Jensen’s inequality, and the second one follows from
and the observation that . In other words, random projections reduce the Lipschitz constant by a factor of on average. Since learning requires samples, this allows us to work with exponentially fewer samples and makes the learning task easier. Conversely, given a fixed training dataset, it gives more accurate estimates.
3.1.1 The case for Delaunay triangulations
The above example uses unstructured random subspaces. In many inverse problems, such as inverse scattering (Beretta et al. [2013], Di Cristo and Rondi [2003]), a judicious choice of subspace family can give exponential improvements in Lipschitz stability. Particularly, it is favorable to use piecewiseconstant images: with being indicator functions of some domain subset.
Motivated by this observation, we use piecewiseconstant subspaces over random Delaunay triangle meshes. The Delaunay triangulations enjoy a number of desirable learningtheoretic properties. For function learning it was shown that given a set of vertices, piecewise linear functions on Delaunay triangulations achieve the smallest supnorm error among all triangulations (Omohundro [1989]).
We sample sets of points in the image domain from a uniformdensity Poisson process and construct (discrete) Delaunay triangulations with those points as vertices. Let be the collection of subspaces of piecewiseconstant functions on these triangulations. Let further be the map from to the projection of the model into subspace , . Instead of learning the “hard” inverse mapping , we propose to learn an ensemble of simpler mappings .
We approximate each by a convolutional neural network, , parameterized by a set of trained weights . Similar to Jin et al. [2016], we do not use the measured data directly as this would require the network to first learn to map back to the image domain; we rather warmstart the reconstruction by a nonnegative least squares reconstruction, , computed from . The weights are chosen by minimizing empirical risk:
(1) 
where is a set of training models and nonnegative least squares measurements.
3.2 The new inverse problem
By learning projections onto random subspaces, we transform our original problem into that of estimating from . To see how this can be done, ascribe to the columns of a natural orthogonal basis for the subspace , with being the indicator function of the th triangle in mesh . Denote by the mapping from the data to an estimate of the expansion coefficients of in the basis for :
Let , and ; then we can estimate using the following reformulated problem:
and the corresponding regularized reconstruction:
(2) 
with chosen as the TVseminorm . The regularization is not essential. As we show experimentally, if is sufficiently large, is not required. Note that solving the original problem directly using regularizer fails to recover the structure of the model (Figure 1).
3.3 Stability of the reformulated problem and “convolutionalization”
Since the true inverse map has a large Lipschitz constant, it would seem reasonable that as the number of mesh subspaces grows large (and their direct sum approaches the whole ambient space ), the Lipschitz properties of should deteriorate as well.
Denote the unregularized inverse mapping in (2) by . Then we have the following estimate:
with
the smallest (nonzero) singular value of
and the Lipschitz constant of the stable projection mappings . Indeed, we observe empirically that grows large as the number of subspaces increases which reflects the fact that although individual projections are easier to learn, the fullresolution reconstruction remains illposed.Estimates of individual subspace projections give correct local information. They convert possibly nonlocal measurements (e.g. integrals along curves in tomography) into local ones. The key is that these local averages (subspace projection coefficients) can be estimated accurately (see Section 4).
To further illustrate what we mean by correct local information, consider a simple numerical experiment with our reformulated problem, , where is an allzero image with a few pixels “on”. For the sake of clarity we assume the coefficients are perfect. Recall that is a block matrix comprising
subspace bases stacked side by side. It is a random matrix because the subspaces are generated at random, and therefore the reconstruction
is also random. We approximate by simulating a large number of tuples of meshes and averaging the obtained reconstructions.Results are shown in Figure 3 for different numbers of triangles per subspace, , and subspaces per reconstruction, . As or increase, the expected reconstruction becomes increasingly localized around nonzero pixels. The following proposition (proved in Appendix B) tells us that this phenomenon can be modeled by convolution.^{4}^{4}4We note that this result requires adequate handling of boundary conditions; for the lack of space we omit the straightforward details.
Let be the solution to given as . Then there exists a kernel , with a discrete index, such that . Furthermore, is isotropic.
While Figure 3 suggests that more triangles are better, we note that this increases the subspace dimension which makes getting correct projection estimates harder. Instead we choose to stack more meshes with a smaller number of triangles.
Intuitively, since every triangle average depends on many measurements, estimating each average is more robust to measurement corruptions as evidenced in Section 4. Accurate estimates of local averages enable us to recover the geometric structure while being more robust to data errors.
4 Numerical results
4.1 Application: traveltime tomography
To demonstrate our method’s benefits we consider linearized traveltime tomography (Hole [1992], Bording et al. [1987]), but we note that the method applies to any inverse problem with scarce data.
In traveltime tomography, we measure wave travel times between sensors as in Figure 4. Travel times depend on the medium property called slowness (inverse of speed) and the task is to reconstruct the spatial slowness map. Image intensities are a proxy for slowness maps—the lower the image intensity the higher the slowness. In the straightray approximation, the problem data is modeled as integral along line segments:
(3) 
where is the continuous slowness map and are sensor locations. In our experiments, we use a pixel grid with sensors (300 measurements) placed uniformly in an inscribed circle, and corrupt the measurements with zeromean iid Gaussian noise.
4.1.1 Architectures and reconstruction
We generate random Delaunay meshes each with 50 triangles. The corresponding projector matrices compute average intensity over triangles to yield a piecewise constant approximation of . We test two distinct architectures: (i) ProjNet, tasked with estimating the projection into a single subspace; and (ii) SubNet, tasked with estimating the projection over multiple subspaces.^{5}^{5}5Code available at https://github.com/swingresearch/deepmesh under the MIT License.
The ProjNet architecture is inspired by the FBPConvNet (Jin et al. [2016]) and the UNet (Ronneberger et al. [2015]) as shown in Figure 11a in the appendix. Crucially, we constrain the network output to live in by fixing the last layer of the network to be a projector, (Figure 11a). A similar trick in a different context was proposed in (Sønderby et al. [2016]).
We combine projection estimates from many ProjNets by regularized linear leastsquares (2) to get the reconstructed model (cf. Figure 2) with the regularization parameter determined on five heldout images. A drawback of this approach is that a separate ProjNet must be trained for each subspace. This motivates the SubNet (shown in Figure 11b). Each input to SubNet is the concatenation of a nonnegative least squares reconstruction and 50 basis functions, one for each triangle forming a 51channel input. This approach scales to any number of subspaces which allows us to get visually smoother reconstructions without any further regularization as in (2). On the other hand, the projections are less precise which can lead to slightly degraded performance.
As a quantitative figure of merit we use the signaltonoise ratio (SNR). The input SNR is defined as where and
are the signal and noise variance; the output SNR is defined as
with the ground truth and the reconstruction.130 ProjNets are trained for 130 different meshes with measurements at various SNRs. Similarly, a single SubNet is trained with 350 different meshes and the same noise levels. We compare the ProjNet and SubNet reconstructions with a direct Unet baseline convolutional neural network that reconstructs images from their nonnegative least squares reconstructions. The direct baseline has the same architecture as SubNet except the input is a single channel nonnegative least squares reconstruction like in ProjNet and the output is the target reconstruction. Such an architecture was proposed by (Jin et al. [2016]) and is used as a baseline in recent learningbased inverse problem works (Lunz et al. [2018], Ye et al. [2018]) and is inspiring other architectures for inverse problems (Antholzer et al. [2017]). We pick the best performing baseline network from multiple networks which have a comparable number of trainable parameters to SubNet. We simulate the lack of training data by testing on a dataset that is different than that used for training.
Robustness to corruption
To demonstrate that our method is robust against arbitrary assumptions made at training time, we consider two experiments. First, we corrupt the data with zeromean iid Gaussian noise and reconstruct with networks trained at different input noise levels. In Figures 5a, 12 and Table 1, we summarize the results with reconstructions of geo images taken from the BP2004 dataset^{6}^{6}6http://software.seg.org/datasets/2D/2004_BP_Vel_Benchmark/ and xray images of metal castings (Mery et al. [2015]). The direct baseline and SubNet are trained on a set of 20,000 images from the arbitrarily chosen LSUN bridges dataset (Yu et al. [2015]) and tested with the geophysics and xray images. ProjNets are trained with 10,000 images from the LSUN dataset. Our method reports better SNRs compared with the baseline. We note that direct reconstruction is unstable when trained on clean and tested on noisy measurements as it often hallucinates details that are artifacts of the training data. For applications in geophysics it is important that our method correctly captures the shape of the cavities unlike the direct inversion which can produce sharp but wrong geometries (see outlines in Figure 5a).

Training SNR  

10 dB  dB  
Direct  ProjNets  SubNet  Direct  ProjNets  SubNet  

10 dB  13.51  14.49  13.92  10.34  12.88  12.85  
dB  13.78  15.38  14.04  16.67  17.23  16.86 
Second, we consider a different corruption mechanism where traveltime measurements are erased (set to zero) independently with probability , and use networks trained with 10 dB input SNR on the LSUN dataset to reconstruct. Figure 5b and Table 2 summarizes our findings. Unlike with Gaussian noise (Figure 5a) the direct method completely fails to recover coarse geometry in all test cases. In our entire test dataset of 102 xray images there is not a single example where the direct network captures a geometric feature that our method misses. This demonstrates the strengths of our approach. For more examples of xray images please see Appendix E.
Robustness against dataset overfitting
Figure 6 illustrates the influence of the training data on reconstructions. Training with LSUN, CelebA (Liu et al. [2015]) and a synthetic dataset of random overlapping shapes (see Figure 15 in Appendix for examples) all give comparable reconstructions—a desirable property in applications where real ground truth is unavailable.
We complement our results with reconstructions of checkerboard phantoms (standard resolution tests) and xrays of metal castings in Figure 7. We note that in addition to better SNR, our method produces more accurate geometry estimates, as per the annotations in the figure.
5 Conclusion
We proposed a new approach to regularize illposed inverse problems in imaging, the key idea being to decompose an unstable inverse mapping into a collection of stable mappings which only estimate lowdimensional projections of the model. By using piecewiseconstant Delaunay subspaces, we showed that the projections can indeed be accurately estimated. Combining the projections leads to a deconvolutionlike problem. Compared to directly learning the inverse map, our method is more robust against noise and corruptions. We also showed that regularizing via projections allows our method to generalize across training datasets. Our reconstructions are better both quantitatively in terms of SNR and qualitatively in the sense that they estimate correct geometric features even when measurements are corrupted in ways not seen at training time. Future work involves getting precise estimates of Lipschitz constants for various inverse problems, regularizing the reformulated problem using modern regularizers (Ulyanov et al. [2017]), studying extensions to nonlinear problems and developing concentration bounds for the equivalent convolution kernel.
Acknowledgement
This work utilizes resources supported by the National Science Foundation’s Major Research Instrumentation program, grant #1725729, as well as the University of Illinois at UrbanaChampaign.
References
 Adler and Öktem [2017a] Jonas Adler and Ozan Öktem. Solving illposed inverse problems using iterative deep neural networks. arXiv preprint arXiv:1704.04058v2, April 2017a.
 Adler and Öktem [2017b] Jonas Adler and Ozan Öktem. Learned Primaldual Reconstruction. arXiv preprint arXiv:1707.06474v1, July 2017b.
 Antholzer et al. [2017] Stephan Antholzer, Markus Haltmeier, and Johannes Schwab. Deep Learning for Photoacoustic Tomography from Sparse Data. arXiv preprint arXiv:1704.04587v2, April 2017.
 ArayaPolo et al. [2017] Mauricio ArayaPolo, Joseph Jennings, Amir Adler, and Taylor Dahlke. Deeplearning tomography. The Leading Edge, December 2017.
 Beretta et al. [2013] Elena Beretta, Maarten V de Hoop, and Lingyun Qiu. Lipschitz Stability of an Inverse Boundary Value Problem for a SchrödingerType Equation. SIAM J. Math. Anal., 45(2):679–699, March 2013.
 Bianco and Gertoft [2017] Michael Bianco and Peter Gertoft. Sparse travel time tomography with adaptive dictionaries. arXiv preprint arXiv:1712.08655, 2017.
 Bora et al. [2017] Ashish Bora, Ajil Jalal, Eric Price, and Alexandros G Dimakis. Compressed sensing using generative models. arXiv preprint arXiv:1703.03208, 2017.
 Bora et al. [2018] Ashish Bora, Eric Price, and Alexandros G Dimakis. Ambientgan: Generative models from lossy measurements. In International Conference on Learning Representations (ICLR), 2018.
 Bording et al. [1987] R Phillip Bording, Adam Gersztenkorn, Larry R Lines, John A Scales, and Sven Treitel. Applications of seismic traveltime tomography. Geophysical Journal International, 90(2):285–303, 1987.
 Cooper [1995] Duane A Cooper. Learning lipschitz functions. International Journal of Computer Mathematics, 59(12):15–26, 1995.
 Di Cristo and Rondi [2003] Michele Di Cristo and Luca Rondi. Examples of exponential instability for inverse inclusion and scattering problems. Inverse Problems, 19(3):685, 2003.
 Galetti et al. [2017] Erica Galetti, Andrew Curtis, Brian Baptie, David Jenkins, and Heather Nicolson. Transdimensional Lovewave tomography of the British Isles and shearvelocity structure of the East Irish Sea Basin from ambientnoise interferometry. Geophys. J. Int., 208(1):36–58, January 2017.
 Greenspan et al. [2016] Hayit Greenspan, Bram van Ginneken, and Ronald M Summers. Deep Learning in Medical Imaging: Overview and Future Promise of an Exciting New Technique. IEEE Trans. Med. Imag., 35(5):1153–1159, may 2016.
 Gregor and LeCun [2010] Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. In Proceedings of the 27th International Conference on International Conference on Machine Learning, pages 399–406. Omnipress, 2010.
 Gribonval et al. [2017] Rémi Gribonval, Gilles Blanchard, Nicolas Keriven, and Yann Traonmilin. Compressive statistical learning with random feature moments. arXiv preprint arXiv:1706.07180, 2017.

Güler and Übeylı [2005]
İnan Güler and Elif Derya Übeylı.
ECG beat classifier designed by combined neural network model.
Pattern Recognition, 38(2):199–208, 2005.  Han et al. [2016] Yo Seob Han, Jaejun Yoo, and Jong Chul Ye. Deep Residual Learning for Compressed Sensing CT Reconstruction via Persistent Homology Analysis. arXiv preprint arXiv:1611.06391, November 2016.

He et al. [2016a]
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Identity mappings in deep residual networks.
In
European Conference on Computer Vision
, pages 630–645. Springer, 2016a.  He et al. [2016b] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition, pages 770–778. IEEE, 2016b.
 Hole [1992] John Hole. Nonlinear highresolution threedimensional seismic travel time tomography. Journal of Geophysical Research: Solid Earth, 97(B5):6553–6562, 1992.
 Hoole [1993] S R H Hoole. Artificial neural networks in the solution of inverse electromagnetic field problems. IEEE Trans. Magn., 29(2):1931–1934, March 1993.

Hudson and Cohen [2000]
Donna L Hudson and Maurice E Cohen.
Neural networks and artificial intelligence for biomedical engineering
. Wiley Online Library, 2000.  Jin et al. [2016] Kyong Hwan Jin, Michael T McCann, Emmanuel Froustey, and Michael Unser. Deep Convolutional Neural Network for Inverse Problems in Imaging. arXiv preprint arXiv:1611.03679v1, November 2016.
 Kelly et al. [2017] Brendan Kelly, Thomas P Matthews, and Mark A Anastasio. Deep LearningGuided Image Reconstruction from Incomplete Data. arXiv preprint arXiv:1709.00584, September 2017.
 Lewis et al. [2017] Winston Lewis, Denes Vigh, et al. Deep learning prior models from seismic images for fullwaveform inversion. In SEG International Exposition and Annual Meeting. Society of Exploration Geophysicists, 2017.
 Li et al. [2018] Housen Li, Johannes Schwab, Stephan Antholzer, and Markus Haltmeier. NETT: Solving Inverse Problems with Deep Neural Networks. arXiv preprint arXiv:1803.00092v1, February 2018.
 Liu et al. [2015] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015.
 Lucas et al. [2018] Alice Lucas, Michael Iliadis, Rafael Molina, and Aggelos K Katsaggelos. Using Deep Neural Networks for Inverse Problems in Imaging: Beyond Analytical Methods. IEEE Signal Process. Mag., 35(1):20–36, 2018.
 Lunz et al. [2018] Sebastian Lunz, Ozan Öktem, and CarolaBibiane Schönlieb. Adversarial regularizers in inverse problems. arXiv preprint arXiv:1805.11572, 2018.
 McCann et al. [2017] Michael T McCann, Kyong Hwan Jin, and Michael Unser. Convolutional neural networks for inverse problems in imaging: A review. IEEE Signal Process. Mag., 34(6):85–95, 2017.
 Mery et al. [2015] Domingo Mery, Vladimir Riffo, Uwe Zscherpel, Germán Mondragón, Iván Lillo, Irene Zuccar, Hans Lobel, and Miguel Carrasco. GDXray: The Database of Xray Images for Nondestructive Testing. Journal of Nondestructive Evaluation, 34, 11 2015.
 Ogawa et al. [1998] Takehiko Ogawa, Yukio Kosugi, and Hajime Kanada. Neural network based solution to inverse problems. In Neural Networks Proceedings, 1998. IEEE World Congress on Computational Intelligence. The 1998 IEEE International Joint Conference on, volume 3, pages 2471–2476. IEEE, 1998.
 Omohundro [1989] S M Omohundro. The Delaunay triangulation and function learning, 1989.
 Pelt and Batenburg [2013] Daniel Maria Pelt and Kees Joost Batenburg. Fast tomographic reconstruction from limited data using artificial neural networks. IEEE Trans. on Image Process., 22(12):5238–5251, 2013.
 Pilanci and Wainwright [2016] Mert Pilanci and Martin J Wainwright. Iterative hessian sketch: Fast and accurate solution approximation for constrained leastsquares. The Journal of Machine Learning Research, 17(1):1842–1879, 2016.
 Porikli et al. [2017] Fatih Porikli, Shiguang Shan, Cees Snoek, Rahul Sukthankar, and Xiaogang Wang. Deep Learning for Visual Understanding [From the Guest Editors]. IEEE Signal Process. Mag., 34(6):24–25, Nov 2017.
 Porikli et al. [2018] Fatih Porikli, Shiguang Shan, Cees Snoek, Rahul Sukthankar, and Xiaogang Wang. Deep Learning for Visual Understanding: Part 2 [From the Guest Editors]. IEEE Signal Process. Mag., 35(1):17–19, Jan 2018.
 Rahimi and Recht [2008] Ali Rahimi and Benjamin Recht. Random features for largescale kernel machines. Advances in Neural Information and Processing (NIPS), 2008.
 Rahimi and Recht [2009] Ali Rahimi and Benjamin Recht. Weighted Sums of Random Kitchen Sinks: Replacing minimization with randomization in learning. Advances in Neural Information and Processing (NIPS), pages 1313–1320, 2009.
 Rick Chang et al. [2017] JH Rick Chang, ChunLiang Li, Barnabas Poczos, BVK Vijaya Kumar, and Aswin C Sankaranarayanan. One Network to Solve Them All–Solving Linear Inverse Problems Using Deep Projection Models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5888–5897, 2017.
 Ronneberger et al. [2015] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. Unet: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and ComputerAssisted Intervention, pages 234–241. Springer, 2015.
 Schiller and Doerffer [2010] Helmut Schiller and Roland Doerffer. Neural network for emulation of an inverse model operational derivation of Case II water properties from MERIS data. International Journal of Remote Sensing, November 2010.
 Sønderby et al. [2016] Casper Kaae Sønderby, Jose Caballero, Lucas Theis, Wenzhe Shi, and Ferenc Huszár. Amortised map inference for image superresolution. arXiv preprint arXiv:1610.04490, 2016.
 Sprechmann et al. [2013] Pablo Sprechmann, Roee Litman, Tal Ben Yakar, Alexander M Bronstein, and Guillermo Sapiro. Supervised sparse analysis and synthesis operators. In Advances in Neural Information Processing Systems, pages 908–916, 2013.
 Stefanov and Uhlmann [2009] Plamen Stefanov and Gunther Uhlmann. Linearizing nonlinear inverse problems and an application to inverse backscattering. Journal of Functional Analysis, 256(9):2842–2866, 2009.
 Ulyanov et al. [2017] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. arXiv preprint arXiv:1711.10925, 2017.
 Venkatakrishnan et al. [2013] Singanallur V Venkatakrishnan, Charles A Bouman, and Brendt Wohlberg. Plugandplay priors for model based reconstruction. In Global Conference on Signal and Information Processing (GlobalSIP), 2013 IEEE, pages 945–948. IEEE, 2013.
 Vetterli et al. [2014] Martin Vetterli, Jelena Kovačević, and Vivek K Goyal. Foundations of signal processing. Cambridge University Press, 2014.
 Wang [2016] Ge Wang. A perspective on deep imaging. IEEE Access, 4:8914–8924, 2016.
 Xin et al. [2016] Bo Xin, Yizhou Wang, Wen Gao, David Wipf, and Baoyuan Wang. Maximal sparsity with deep networks? In Advances in Neural Information Processing Systems, pages 4340–4348, 2016.
 Yao et al. [2006] Huajian Yao, Robert D van Der Hilst, and Maarten V De Hoop. Surfacewave array tomography in se tibet from ambient seismic noise and twostation analysis—i. phase velocity maps. Geophysical Journal International, 166(2):732–744, 2006.
 Ye et al. [2018] Jong Chul Ye, Yoseob Han, and Eunju Cha. Deep convolutional framelets: A general deep learning framework for inverse problems. SIAM Journal on Imaging Sciences, 11(2):991–1048, 2018.
 Yu et al. [2015] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. LSUN: Construction of a Largescale Image Dataset using Deep Learning with Humans in the Loop. arXiv preprint arXiv:1506.03365, 2015.
 Yurtsever et al. [2017] Alp Yurtsever, Madeleine Udell, Joel A Tropp, and Volkan Cevher. Sketchy decisions: Convex lowrank matrix optimization with optimal storage. arXiv preprint arXiv:1702.06838, 2017.
 Zhang et al. [2016] Hanming Zhang, Liang Li, Kai Qiao, Linyuan Wang, Bin Yan, Lei Li, and Guoen Hu. Image Prediction for Limitedangle Tomography via Deep Learning with Convolutional Neural Network. arXiv preprint arXiv:1607.08707v1, July 2016.
 Zhu et al. [2018] Bo Zhu, Jeremiah Z Liu, Stephen F Cauley, Bruce R Rosen, and Matthew S Rosen. Image reconstruction by domaintransform manifold learning. Nature, 555(7697):487, March 2018.
Appendix A Need for nonlinear operators
We explain the need for nonlinear operators even in the absence of noise with reference to Figure 8. Projecting into a given known subspace is a simple linear operation, so it may not be a priori clear why we use nonlinear neural networks to estimate the projections. Alas, we do not know and only have access to . Suppose that there exists a linear operator (a matrix) which acts on and computes the projection of on . A natural requirement on is consistency: if already lives in , then we would like to have . This implies that for any , not necessarily in , we require which implies that is an idempotent operator. Letting the columns of be a basis for , it is easy to see that the least squares minimizer for is . However, because ( is the adjoint of , simply a transpose for real matrices), in general it will not hold that . Thus, is an oblique, rather than orthogonal projection into . In Figure 8 this corresponds to the point which can be arbitrarily far from the orthogonal projection . The nullspace of this projection is precisely .
Thus consistent linear operators can at best yield oblique projections which can be far from the orthogonal one. One could also see this geometrically from Figure 8. As the angle between and increases to the oblique projection point travels to infinity (note that the oblique projection always happens along the nullspace of , which is the line orthogonal to . Since our subspaces are chosen at random, in general they are not aligned with . The only subspace on which we can linearly compute an orthogonal projection from is ; this is given by the MoorePenrose pseudoinverse. Therefore, to get the orthogonal projection onto random subspaces, we must use nonlinear operators. More generally, for any other ad hoc linear reconstruction operator , always lives in the column space of which is a subspace whose dimension is at most the number of rows of . However, we do not have any linear subspace model for .
As shown in the right half of Figure 8, as soon as is injective on , the existence of this nonlinear map is guaranteed by construction: since determines , it also determines .
We show the results of numerical experiments in Figures 9 and 10 which further illustrate the performance difference between linear oblique projectors and our nonlinear learned operator when estimating the projection of an image into a random subspace. We refer the reader to the captions below each figure for more details.
Appendix B Proof of proposition 3.3
Proof.
The reconstruction of the new inverse problem can be written as where the columns of form a biorthogonal basis to the columns of . Thus
Using the definition of the inner product and rearranging, we get
where
. Now, the probability distribution of triangles around any point
is both shift and rotationinvariant because a Poisson process in the plane is shift and rotationinvariant. It follows that for some , meaning thatwhich is a convolution of the original model with a rotationally invariant (isotropic) kernel. ∎
Appendix C Network architectures
Figure 11 explains the network architecture used for ProjNet and SubNet. The network consists of a sequence of downsampling layers followed by upsampling layers, with skip connections (He et al. [2016b, a]) between the downsampling and upsampling layers. Each ProjNet output is constrained to a single subspace by applying a subspace projection operator, . We train such networks and reconstruct from the projection estimate using (2). SubNet is a single network that is trained over multiple subspaces. To do this, we change its input to be . Moreover, we apply the same projection operator as ProjNet to the output of the SubNet. Each SubNet is trained to give projection estimates over random subspaces. This approach allows us to scale to any number of subspaces without training new networks for each. Moreover, this allows us to build an overconstrained system to solve. Even though SubNet has almost as many parameters as the direct net, reconstructing via the projection estimates allows SubNet to get higher SNR and more importantly, get better estimates of the coarse geometry than the direct inversion. All networks are trained with the Adam optimizer.
Appendix D Further reconstructions
We showcase more reconstructions on actual geophysics images taken from the BP2004 dataset in Figure 12. Note that all networks were trained on the LSUN bridges dataset.
Appendix E Erasure reconstructions
We show additional reconstructions for the largest corruption case, , for xray images (Figure 13) and geo images (Figure 14). Our method consistently has better SNR. More importantly we note that there is not a single instance where the direct reconstruction gets a feature that our methods do not. The majority of times, the direct network misses a feature of the image. This is highly undesirable in settings such as geophysical imaging.
Comments
There are no comments yet.