PengStarobinets

Back to Psych 221 Projects 2013

Background

The performance of long-distance imaging systems can often be strongly affected by atmospheric turbulence caused by variation of refractive index along the optical transmission path. Such turbulence can produce geometric distortion, space and time-variant defocus blur, and motion blur. An example is shown in the following video of the moon.

Below is the example of atmospheric turbulence that was used for this project:

Moon with turbulence

Recording of the moon with effects of turbulence¹

There have been many approaches to solving this problem that attempt to restore a single high-quality image from an observed frame sequence distorted by air turbulence. As in the video, these approaches, and the approach addressed in this paper, work under the assumption that the scene and the image sensor are both static and that observed motions are due to the air turbulence alone.

The imaging process can be modeled as:

$g_{k}=H_{k}F_{k}f+n_{k}$ where $f$ denotes the ideal image, $F_{k}$ and $H_{k}$ represent the geometric deformation and blurring matrices respectively, $n_{k}$ denotes additive noise, and $g_{k}$ is the k-th observed frame. The key then becomes to basically reverse this process so that we can find the desired corrected image.

Methods

Existing restoration algorithms for this problem can generally be categorized in two ways.

Multi-Frame Reconstruction Framework

First, a nonrigid image registration technique is used to register each observed frame with respect to a fixed reference grid (usually generated from an average of the image sequence to eliminate spatial variation)¹. Next, the registration parameters are used to estimate the corresponding motion field for each frame and one frame is transformed back to a non-deformed position. Finally, a sharp image is formed through a Bayesian reconstruction filter.

Lucky Imaging

Image selection and fusion methods are used to reduce the blurring effects caused by turbulence⁵. The image selection algorithm attempts to find frames of the best quality (lucky frames) from the image stream. The output image is then produced by fusing these best-quality images together. The principle behind this method is that for short-exposuer images, turbulence creates 'mutations' in image quality and randomly makes some images sharp. Lucky-image-based weighting schemes can be used to select the sharpest image or images with certain sharp regions, and combine them to create an image that is sharp everywhere.

Our Method

To correct geometric distortion and reduce space and time-varying blur, we are proposing a new approach in this work capable of restoring a single high-quality image from a given image sequence distorted by atmospheric turbulence. This approach uses two major steps. For the first step, we use a B-spline based non-rigid image registration algorithm to register each observed frame with respect to a reference image^1,4. To improve registration accuracy, a symmetry constraint is introduced to the cost function to penalize inconsistency between the forward and backward deformation parameters during the estimation process. A fast Gauss-Newton method is used to reduce the computational cost of the registration algorithm.

Our new idea and an addition to this approach involves separating a single frame into a set of overlapping sub-frames and running the earlier-described algorithm on those sub-frames. In addition to reducing computational load and time, this improves on the original method by reducing the effect of systematic errors on the frame, since the errors are localized and isolated to the particular block rather than affecting the whole image. Another benefit of this approach is that scaling the number of the test points causes a much smaller processing time increase as opposed to the processing time that would be required for the same number of test points if applied to the whole image. The second step of our proposal is to stitch the undeformed back into a single image and then, using a high-pass sharpening filter produce a final, high-quality image.

Proposed Approach

Non-rigid Image Registration

First, we obtain a reference image without image deformation, generally by averaging the frame sequence. The math is specified below in the equations section. However, in general, the geometric distortion between an observed distorted image and the reference image is obtained by the movement of control points placed in a grid on the images. B-spline interpolation is then used to map all the points from the reference image to the deformed image¹. A cost function is then calculated to minimize both the error between the mapped points and maximize symmetry between forward and backward deformations. After this cost function is minimized, a motion vector from the deformed image to the reference image is generated. This process is repeated for each sub-frame and then the sub-frames are recombined into a single image.

The figure below shows an example of the sub-frame implementation of turbulence compensation.

(Left) Test points and their deformed positions for each sub-frame. (Right) Deformed location of each pixel for a single subframe.

Equations

Equations analyzed and used are from [1] and [3]. The initial test points are taken with equal spacing and can be represented as: ${\hat {\mathbf {x} }}_{0i}=({\hat {x}}_{0i},{\hat {y}}_{0i})^{T}$

Running correlation algorithm, we found the deformed locations of the test points. This difference between the original positions and the deformed position are then stored in the deformation vector:

${\vec {\mathbf {p} }}=[{\hat {x}}_{0i}-{\hat {x}}_{i},{\hat {y}}_{0i}-{\hat {y}}_{i}]^{T}$ = $[\Delta {\hat {x}}_{1},...,\Delta {\hat {x}}_{i},\Delta {\hat {y}}_{1},...,\Delta {\hat {y}}_{i}]^{T}$

Using B-spline interpolation, we defined the spline basis $c_{i}$ and the basis function matrix $\mathbf {A{(x)}}$ for each pixel where $\mathbf {x} =(x,y)^{T}$.

By definition, $c_{i}$ is:

$c_{i}=\mathrm {B} ({\frac {x-{\hat {x}}_{0i}}{h_{x}}})$ $\mathrm {B} ({\frac {y-{\hat {y}}_{0i}}{h_{y}}})$, where $\mathrm {B} (z)={\begin{cases}2/3-(1-\left|z\right|/2)z^{2}{\mbox{, if }}0\leq \left|z\right|\leq 1\\(2-\left|z\right|)^{3}/6{\mbox{, if }}1\leq \left|z\right|\leq 2\\0{\mbox{, otherwise}}\\\end{cases}}$

The basis function matrix is then:

$\mathbf {A(x)} ={\begin{bmatrix}c_{1}&\cdots &c_{n}&0&\cdots &0\\0&\cdots &0&c_{1}&\cdots &c_{n}\end{bmatrix}}$

With the basis function matrix, we can then define the deformed coordinates for every pixel:

$\mathbf {W(x;} {\vec {\mathbf {p} }}\mathbf {)} =\mathbf {x} +\mathbf {A(x)} {\vec {\mathbf {p} }}$

As suggested in Professor Milanfar's paper, instead of using a classic B-spline registration approach to the estimation of the deformation vector through minimization of the following cost function,

$C({\overrightarrow {\mathbf {p} }})=\sum _{x}\left|R(\mathbf {W(x;} {\overrightarrow {\mathbf {p} }}\mathbf {)} )-G(\mathbf {x} )\right|^{2}$

which is prone to local minima traps, one can use the following cost function:

$C({\overrightarrow {\mathbf {p} }})=\sum _{x}\left|R(\mathbf {W(x;} {\overrightarrow {\mathbf {p} }}\mathbf {)} )-G(\mathbf {x} )\right|^{2}+\sum _{x}\left|G(\mathbf {W(x;} {\overleftarrow {\mathbf {p} }}\mathbf {)} )-R(\mathbf {x} )\right|^{2}+\gamma ({\overrightarrow {\mathbf {p} }}+{\overleftarrow {\mathbf {p} }})^{T}({\overrightarrow {\mathbf {p} }}+{\overleftarrow {\mathbf {p} }})$

by initial approximation that:

$\mathbf {p^{T}} =[{\overrightarrow {\mathbf {p} }}^{T},{\overleftarrow {\mathbf {p} }}^{T}]$

where we initially assume that

${\overleftarrow {\mathbf {p} }}=-{\overrightarrow {\mathbf {p} }}$

Then, after each iteration, the cost deformation vector can be updated as:

$\mathbf {p} ^{l+1}=\mathbf {p} ^{l}-\mathbf {H} ^{-1}\mathbf {b}$

The following equations describe the system:

$\mathbf {H} ={\begin{bmatrix}{\overrightarrow {\mathbf {H} }}+\mathbf {\gamma I} &\mathbf {\gamma I} \\\mathbf {\gamma I} &{\overleftarrow {\mathbf {H} }}+\mathbf {\gamma I} \end{bmatrix}}$, where $\gamma =5000$. ${\overrightarrow {\mathbf {H} }}=\sum _{x}{\overrightarrow {\mathbf {d} }}\mathbf {(x)} {\overrightarrow {\mathbf {d} }}\mathbf {(x)} ^{T},{\overleftarrow {\mathbf {H} }}=\sum _{x}{\overleftarrow {\mathbf {d} }}\mathbf {(x)} {\overleftarrow {\mathbf {d} }}\mathbf {(x)} ^{T}$

The faster implementation, as suggested in the paper is calculating ${\overrightarrow {\mathbf {H} }}$ and ${\overleftarrow {\mathbf {H} }}$ the following way:

${\overrightarrow {\mathbf {H} }}=\sum _{x}{\overrightarrow {\mathbf {H_{x}} }}$

where:

${\overrightarrow {\mathbf {H_{x}} }}={\begin{bmatrix}r_{x}(\mathbf {W} )^{2}\mathbf {C_{x}} &r_{x}(\mathbf {W} )r_{y}(\mathbf {W} )\mathbf {C_{x}} \\r_{x}(\mathbf {W} )r_{y}(\mathbf {W} )\mathbf {C_{x}} &r_{y}(\mathbf {W} )^{2}\mathbf {C_{x}} \end{bmatrix}}$ ${\overrightarrow {\mathbf {d} }}\mathbf {(x)^{T}} ={\frac {\partial R(\mathbf {W(x;} {\overrightarrow {\mathbf {p} }}^{l}\mathbf {)} )}{\partial \mathbf {W} }}\mathbf {A(x)} ,{\overleftarrow {\mathbf {d} }}\mathbf {(x)^{T}} ={\frac {\partial G(\mathbf {W(x;} {\overleftarrow {\mathbf {p} }}^{l}\mathbf {)} )}{\partial \mathbf {W} }}\mathbf {A(x)}$ ${\overrightarrow {\mathbf {b} }}=\sum _{x}{\overrightarrow {\mathbf {d} }}\mathbf {(x)} [R(\mathbf {W(x;} {\overrightarrow {\mathbf {p} }}^{l}))-G(\mathbf {x} )],{\overleftarrow {\mathbf {b} }}=\sum _{x}{\overleftarrow {\mathbf {d} }}\mathbf {(x)} [G(\mathbf {W(x;} {\overleftarrow {\mathbf {p} }}^{l}))-R(\mathbf {x} )]$

where ${\frac {\partial R(\mathbf {W(x;} {\overrightarrow {\mathbf {p} }}^{l}\mathbf {)} )}{\partial \mathbf {W} }}=[r_{x}(\mathbf {W} ),r_{y}(\mathbf {W} )]$ and is the gradient of the averaged image R at the deformed pixel location $\mathbf {W(x;} {\overrightarrow {\mathbf {p} }}^{l}\mathbf {)} )$.

Results

After a number of iterations to minimize the cost function, the new deformation matrix $\mathbf {W(x;} {\overleftarrow {\mathbf {p} }})$ for a single subframe can look like this:

Deformation matrix after 10 iterations

Using our process flow, we generated a set of sub-frames. As can be seen in the figure below, each individual sub-frame suffers from edge artifacts. Overlapping removes these artifacts and the final stitched image is also shown.

Frame before and after the stitching algorithm

The images below show the frames as the progression from distorted image to the final improved one.

Process Flow

Conclusions

In this project, we proposed a new framework for restoration of a single high-quality frame from an image sequence distorted by air turbulence. The proposed algorithm uses a multi-frame reconstruction approach with the addition of sub-frame evaluations with overlap and sub-image stitching to create a final image. The algorithm registers the sub-frames to suppress spatial and geometric deformation using a B-spline based non-rigid image registration method. Next the frames are stitched together using an edge-correlation algorithm. Finally, a sharpening high-pass filter is applied to the final image to generate a final high-quality output. We feel that our results show that our method is both an effective implementation of the multi-frame reconstruction framework and possibly improves on the process in some ways, by preventing localized errors from affecting the whole image.

References

[1] Zhu, X. and Milanfar, P., “Image Reconstruction from Videos Distorted by Atmospheric Turbulance”, SPIE Electronic Imaging, Conference 7543 on Visual Information Processing and Communication, San Jose, CA, 2010

[2] Shimizu, M., Yoshimura, S., Tanaka, M., “Super-Resolution from Image Sequence under Influence of Hot-Air Optical Turbulence”, CVPR 2008 , 1–8 (June 2008)

[3] Zhu, X. and Milanfar P., “Removing Atmospheric Turbulence via Space-Invariant Deconvolution”, IEEE Trans. on Pattern Analysis and Machine Intelligence vol. 35, no. 1, pp. 157-170, Jan. 2013

[4] Szelinski, R. and Coughlan J., “Spline-Based Image Registration”, CRL, Jan. 1994

[5] Joshi, N. and Cohen M., “Seeing Mt. Rainier: Lucky Imaging for Multi-Image Denoising, Sharpening, and Haze Removal”, Microsoft Research, ICCP, 2010

Software

This project has been done using MATLAB R2011b.

Appendix I - Code and Data

Code

File:Code JPTS.zip

Data

The data has been provided by Professor Milanfar. Originally provided by NASA Langley Research Center.

MoonSurface data

Appendix II - Work partition

Both of us have been working together on the project/wiki as a team with even work distribution.