A portion of the disclosure of this patent document contains material that is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever.
The present invention relates generally to electronic imaging, and particularly to reconstruction of high-quality images from large volumes of low-quality image data.
A number of authors have proposed image sensors with dense arrays of one-bit sensor elements (also referred to as “jots” or binary pixels). The pitch of the sensor elements in the array can be less than the optical diffraction limit. Such binary sensor arrays can be considered a digital emulation of silver halide photographic film. This idea has been recently implemented, for example, in the “Gigavision” camera developed at the Ecole Polytechnique Fédérale de Lausanne (Switzerland).
As another example, U.S. Patent Application Publication 2014/0054446, whose disclosure is incorporated herein by reference, describes an integrated-circuit image sensor that includes an array of pixel regions composed of binary pixel circuits. Each binary pixel circuit includes a binary amplifier having an input and an output. The binary amplifier generates a binary signal at the output in response to whether an input voltage at the input exceeds a switching threshold voltage level of the binary amplifier.
Embodiments of the present invention that are described hereinbelow provide improved methods, apparatus and software for image reconstruction from low-quality input.
There is therefore provided, in accordance with an embodiment of the invention, a method for image reconstruction, which includes defining a dictionary including a set of atoms selected such that patches of natural images can be represented as linear combinations of the atoms. A binary input image, including a single bit of input image data per input pixel, is captured using an image sensor. A maximum-likelihood (ML) estimator is applied, subject to a sparse synthesis prior derived from the dictionary, to the input image data so as to reconstruct an output image including multiple bits per output pixel of output image data.
In a disclosed embodiment, capturing the binary input image includes forming an optical image on the image sensor using objective optics with a given diffraction limit, while the image sensor includes an array of sensor elements with a pitch finer than the diffraction limit. Additionally or alternatively, capturing the binary input image includes comparing the accumulated charge in each input pixel to a predetermined threshold, wherein the accumulated charge in each input pixel in any given time frame follows a Poisson probability distribution.
Typically, defining the dictionary includes training the dictionary over a collection of natural image patches so as to find the set of the atoms that best represents the image patches subject to a sparsity constraint.
In a disclosed embodiment, applying the ML estimator includes applying the ML estimator, subject to the sparse synthesis prior, to each of a plurality of overlapping patches of the binary input image so as to generate corresponding output image patches, and pooling the output image patches to generate the output image.
In some embodiments, applying the ML estimator includes applying an iterative shrinkage-thresholding algorithm (ISTA), subject to the sparse synthesis prior, to the input image data. In one embodiment, applying the ISTA includes training a feed-forward neural network to perform an approximation of the ISTA, and applying the ML estimator includes generating the output image data using the neural network.
Additionally or alternatively, applying the ML estimator includes training a feed-forward neural network to perform an approximation of an iterative ML solution, subject to the sparse synthesis prior, and applying the ML estimator includes inputting the input image data to the neural network and receiving the output image data from the neural network. In a disclosed embodiment, the neural network includes a sequence of layers, wherein each layer corresponds to an iteration of the iterative ML solution. Additionally or alternatively, training the feed-forward neural network includes initializing parameters of the neural network based on the iterative ML solution, and then refining the neural network in an iterative adaptation process using the library.
There is also provided, in accordance with an embodiment of the invention, apparatus for image reconstruction, including a memory, which is configured to store a dictionary including a set of atoms selected such that patches of natural images can be represented as linear combinations of the atoms. A processor is configured to receive a binary input image, including a single bit of input image data per pixel, captured by an image sensor, and to apply a maximum-likelihood (ML) estimator, subject to a sparse synthesis prior derived from the dictionary, to the input image data so as to reconstruct an output image including multiple bits per pixel of output image data.
There is additionally provided, in accordance with an embodiment of the invention, a computer software product, including a computer-readable medium in which program instructions are stored, which instructions, when read by a computer, cause the computer to access a dictionary including a set of atoms selected such that patches of natural images can be represented as linear combinations of the atoms, to receive a binary input image, including a single bit of input image data per pixel, captured by an image sensor, and to apply a maximum-likelihood (ML) estimator, subject to a sparse synthesis prior derived from the dictionary, to the input image data so as to reconstruct an output image including multiple bits per pixel of output image data.
There is further provided, in accordance with an embodiment of the invention, apparatus for image reconstruction, including an interface and a processor, which is configured to access, via the interface, a dictionary including a set of atoms selected such that patches of natural images can be represented as linear combinations of the atoms, to receive a binary input image, including a single bit of input image data per pixel, captured by an image sensor, and to apply a maximum-likelihood (ML) estimator, subject to a sparse synthesis prior derived from the dictionary, to the input image data so as to reconstruct an output image including multiple bits per pixel of output image data.
The present invention will be more fully understood from the following detailed description of the embodiments thereof, taken together with the drawings in which:
Dense, binary sensor arrays can, in principle, mimic the high resolution and high dynamic range of photographic films. A major bottleneck in the design of electronic imaging systems based on such sensors is the image reconstruction process, which is aimed at producing an output image with high dynamic range from the spatially-oversampled binary measurements provided by the sensor elements. Each sensor element receives a very low photon count, which is physically governed by Poisson statistics. The extreme quantization of the Poisson statistics is incompatible with the assumptions of most standard image processing and enhancement frameworks. An image processing approach based on maximum-likelihood (ML) approximation of pixel intensity values can, in principle, overcome this difficulty, but conventional ML approaches to image reconstruction from binary input pixels still suffer from image artifacts and high computational complexity.
Embodiments of the present invention that are described herein provide novel techniques that resolve the shortcomings of the ML approach and can thus reconstruct high-quality output images (with multiple bits per output pixel) from binary input image data (comprising a single bit per input pixel) with reduced computational effort. The disclosed embodiments apply a reconstruction algorithm to binary input images using an inverse operator that combines an ML data fitting term with a synthesis term based on a sparse prior probability distribution, commonly referred to simply as a “sparse prior.” The sparse prior is derived from a dictionary, which is trained in advance, for example using a collection of natural image patches. The reconstruction computation is typically applied to overlapping patches in the input binary image, and the patch-by-patch results are then pooled together to generate the reconstructed output image.
In some embodiments, the image reconstruction is performed by applying an iterative shrinkage-thresholding algorithm (ISTA) (possibly of the fast iterative shrinkage-thresholding algorithm (FISTA) type) in order to carry out the ML estimation. Additionally or alternatively, a neural network can be trained to perform an approximation of the ISTA (or FISTA) fitting process, with a small, predetermined number of iterations, or even only a single iteration, and thus to implement an efficient, hardware-friendly, real-time approximation of the inverse operator. The neural network can output results patch-by-patch, or it can be trained to carry out the pooling stage of the reconstruction process, as well.
The methods and apparatus for image reconstruction that are described herein can be useful, inter alia, in producing low-cost consumer cameras based on high-density sensors that output low-quality image data. As another example, embodiments of the present invention may be applied in medical imaging systems, as well as in other applications in which image input is governed by highly-quantized Poisson statistics, particularly when reconstruction throughput is an issue.
Image sensor 26 outputs a binary raw image 30, which is characterized by low dynamic range (one bit per pixel) and high spatial density, with a pixel pitch that is finer than the diffraction limit of optics 24. An ML processor 34 processes image 30, using a sparse prior that is stored in a memory 32, in order to generate an output image 36 with high dynamic range and low noise. Typically, the sparse prior is based on a dictionary D stored in the memory, as explained further hereinbelow.
To model the operation of system 20, we denote by the matrix x the radiant exposure at the aperture of camera 22 measured over a given time interval. This exposure is subsequently degraded by the optical point spread function of optics 24, denoted by the operator H, producing the radiant exposure on image sensor 26: λ=Hx. The number of photoelectrons ejk generated at input pixel j in time frame k follows the Poisson probability distribution with the rate λj, given by:
The binary sensor elements of image sensor 26 compare the accumulated charge against a threshold qi and output a one-bit measurement bjk. Thus, the probability of a given binary pixel j to assume an “off” value in frame k is:
p
j
=P(bjk=0|qj,λj)=P(ejk<qj|qj,λj); (2)
This equation can be written as:
P(bjk|qj,λj)=(1−bjk)pj+bjk(1−pj). (3)
Assuming independent measurements, the negative log likelihood of the radiant exposure x, given the measurements bjk in a binary image B, can be expressed as:
Processor 34 reconstructs output image 36 by solving equation (4), subject to the sparse spatial prior given by the dictionary D. Details of the solution process are described hereinbelow with reference to
In some embodiments, processor 34 comprises a programmable, general-purpose computer processor, which is programmed in software to carry out the functions that are described herein. Memory 32, which holds the dictionary, may be a component of the same computer, and is accessed by processor 34 in carrying out the present methods. Alternatively or additionally, processor 34 may access the dictionary via a suitable interface, such as a computer bus interface or a network interface controller, through which the processor can access the dictionary via a network. The software for carrying out the functions described herein may be downloaded to processor 34 in electronic form, over a network, for example. Additionally or alternatively, the software may be stored on tangible, non-transitory computer-readable media, such as optical, magnetic, or electronic memory media. Further additionally or alternatively, at least some of the functions of processor may be carried out by hard-wired or programmable hardware logic, such as a programmable gate array. An implementation of this latter sort is described in detail in the above-mentioned provisional patent application.
As a preliminary step, processor 34 (or another computer) defines dictionary D, based on a library of known image patches, at a dictionary construction step 40. The dictionary comprises a set of atoms selected such that patches of natural images can be represented as linear combinations of the atoms. The dictionary is constructed by training over a collection of natural image patches so as to find the set of the atoms that best represents the image patches subject to a sparsity constraint.
Processor 34 may access a dictionary that has been constructed and stored in advance, or the processor may itself construct the dictionary at step 40. Techniques of singular value decomposition (SVD) that are known in the art may be used for this purpose. In particular, the inventors have obtained good results in dictionary construction using the k-SVD algorithm described by Aharon et al., in “K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation,” IEEE Transactions on Signal Processing 54(11), pages 4311-4322 (2006), which is incorporated herein by reference. Given a set of signals, such as image patches, K-SVD tries to extract the best dictionary that can sparsely represent those signals. An implementation of K-SVD that can be run for this purpose on the well-known MATLAB toolbox is listed hereinbelow in an Appendix, which is an integral part of the present patent application. K-SVD software is available for download from the Technion Computer Science Web site at the address www.cs.technion.ac.il/˜elad/Various/KSVD_Matlab_ToolBox.zip.
Camera 22 captures a binary image 30 (B) and inputs the image to processor 34, at an image input step 42. Processor 34 now applies ML estimation, using a sparse prior based on the dictionary D, to reconstruct overlapping patches of output image 36 from corresponding patches of the input image, at an image reconstruction step 44. This reconstruction assumes that the radiant exposure λ can be expressed in terms of D by the kernelized sparse representation: λ=Hρ(Dz), wherein z is a vector of coefficients, and ρ is an element-wise intensity transformation function. As one example, for image reconstruction subject to the Poisson statistics of equation (1), the inventors have found a hybrid exponential-linear function to give good results:
wherein c is a constant. Alternatively, other suitable functional representations of ρ may be used.
Processor 34 reconstructs the radiant exposure x at step 44 using the estimator {circumflex over (x)}=ρ(D{circumflex over (z)}), wherein:
The first term on the right-hand side of this equation is the negative log-likelihood fitting term for ML estimation, while ∥z∥1 denotes the l1 norm of the coefficient vector z, which drives the ML solution toward the sparse synthesis prior. The fitting parameter μ can be set to any suitable value, for example μ=4.
In some embodiments, processor 34 solves equation (6) using an iterative optimization algorithm, such as an iterative shrinkage thresholding algorithm (ISTA), or particularly its accelerated version, FISTA, as described by Beck and Teboulle in “A fast iterative shrinkage thresholding algorithm for linear inverse problems,” SIAM Journal on Imaging Sciences 2(1), pages 183-202 (2009), which is incorporated herein by reference. This algorithm is presented below in Listing I, in which σθ is the coordinate-wise shrinking function, with threshold θ and step size η, and the gradient of the negative log-likelihood computed at each iteration is given by:
Using the techniques described above, processor 34 solves equation (6) for each patch of the input binary image B and thus recovers the estimated intensity distribution {circumflex over (x)} of the patch at step 44. Processor 34 pools these patches to generate output image 36, at a pooling step 46. For example, overlapping patches may be averaged together in order to give a smooth output image.
Although the iterative method of solution that is presented above is capable of reconstructing output images with high fidelity (with a substantially higher ratio of peak signal to noise, PSNR, and better image quality than ML estimation alone), the solution can require hundreds of iterations to converge. Furthermore, the number of iterations required to converge to an output image of sufficient quality can vary from image to image. This sort of performance is inadequate for real-time applications, in which fixed computation time is generally required. To overcome this limitation, in an alternative embodiment of the present invention, a small number T of ISTA iterations are unrolled into a feedforward neural network, which subsequently undergoes supervised training on typical inputs for a given cost function f.
z
t+1=σθ(zt−Wdiag(ρ′(Qzt))HT∇l(Hρ(Azt)|B)) (8)
wherein A=Q=D, W=ηDT, and θ=μη1. Each layer 52 corresponds to one such iteration, parameterized by A, Q, W, and θ, accepting zt as input and producing zt+1 as output.
The output of the final layer gives the coefficient vector {circumflex over (z)}=zT, which is then multiplied by the dictionary matrix D, in a multiplier 54, and converted to the radiant intensity {circumflex over (x)}=ρ(D{circumflex over (z)}) by a transformation operator 56.
Layers 52 of neural network 50 are trained by initializing the network parameters as prescribed by equation (8) and then refining the network in an iterative adaptation process, using a training set of N known image patches and their corresponding binary images. The adaptation process can use a stochastic gradient approach, which is set to minimize the reconstruction error F of the entire network, as given by:
Here xn* are the ground truth image patches, and {circumflex over (z)}T(Bn) denotes the output of network 50 with T layers 52, given the binary images Bn corresponding to xn* as input. For a large enough training set, F approximates the expected value of the cost function f corresponding to the standard squared error:
f=1/2∥xn*−ρ(DzT(Bn))∥22. (10)
The output of network 50 and the derivative of the loss F with respect to the network parameters are calculated using forward and back propagation, as summarized in Listings II and III below, respectively. In Listing III, the gradient of the scalar loss F with respect to each network parameter * is denoted by δ*. The gradient with respect to D, δD, is calculated separately, as it depends only on the last iteration of the network.
, outputs of 2: {zt}t=0T, {bt}t=1T
The inventors found that the above training process makes it possible to reduce the number of iterations required to reconstruct {circumflex over (x)} by about two orders of magnitude while still achieving a reconstruction quality comparable to that of ISTA or FISTA. For example, in one experiment, the inventors found that network 50 with only four trained layers 52 was able to reconstruct images with PSNR in excess of 27 dB, while FISTA required about 200 iterations to achieve the same reconstructed image quality. This and other experiments are described in the above-mentioned provisional patent application.
Although the systems and techniques described herein focus specifically on processing of binary images, the principles of the present invention may be applied, mutatis mutandis, to other sorts of low-quality image data, such as input images comprising two or three bits per input pixel, as well as image denoising and low-light imaging, image reconstruction from compressed samples, reconstruction of sharp images over an extended depth of field (EDOF), inpainting, resolution enhancement (super-resolution), and reconstruction of image sequences using discrete event data. Techniques for processing these sorts of low-quality image data are described in the above-mentioned U.S. Provisional Patent Application 62/308,898 and are considered to be within the scope of the present invention.
The work leading to this invention has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement no. 335491.
It will be appreciated that the embodiments described above are cited by way of example, and that the present invention is not limited to what has been particularly shown and described hereinabove. Rather, the scope of the present invention includes both combinations and subcombinations of the various features described hereinabove, as well as variations and modifications thereof which would occur to persons skilled in the art upon reading the foregoing description and which are not disclosed in the prior art.
This application claims the benefit of U.S. Provisional Patent Application 62/308,898, filed Mar. 16, 2016, which is incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
62308898 | Mar 2016 | US |