Anda di halaman 1dari 8

Deblurring Images via Partial Differential Equations

Sirisha L. Kala
Mississippi State University
slk3@msstate.edu
Advisor: Seth F. Oppenheimer

Abstract:

Image deblurring is one of the fundamental problems in the field of image processing and
computer vision. Deblurring an image and sharpening the features within the image
improves the detail and consequently the utility of an image. In this paper we present a
PDE based model to solve the problem using numerical techniques. We restrict ourselves
to one dimensional case. The blurred image is obtained by convolving the original image
with the Gaussian-like point spread function. Deblurring can be treated as running the
diffusion equation backwards in time, with the blurred image as the input. We formulate
the problem as a linear PDE and look for solution operators to get back the desired
original image and analyze the graphical results to evaluate the performance of the
technique. The backward diffusion problem is a very ill-posed problem and directly
solving the inverse equation can lead to very bad results. We will use the method of
quasireversibility to stabilize our numerical approach. Using the software Matlab, we
present animations of the deblurring process with varying parameters.
1. Introduction

Image deblurring is one of the fundamental problems in the field of image processing and
computer vision. Deblurring an image and sharpening the features within the image
improves the detail and consequently the utility of an image. In the field of image
processing one often needs to enhance an image that was blurred or smoothed by some
known operator. It is possible, in some cases, to invert the blurring operator, thereby
deblurring or reconstructing the original image. The principal difficulty in carrying out
this approach is the notoriously ill-conditioned matrices obtained while trying to solve the
unstable backward diffusion problem. This requires some technique of smoothing to
obtain a reasonable solution. It is such an approach which we present in this paper. We
restrict our study to 3 different discrete images on the finite domain [0,1] .

2. Problem Formulation

Given an image or true scene intensity s (x) , the system introduces some amount of
degradation. This is often modeled as a Gaussian blurring operation. The blurred image
or image intensity is obtained by convolving the true scene intensity s(x) with a point
spread function h.


P( x) = ∫ h(u ) s( x − u )du …………………… (1)
−∞

The point spread function in the form of a Gaussian is given by

1  − u2 
h(u ) = exp 
4σπ  4σ 

See reference [1]. The values of P are usually known at only a finite number of points.
These points are called pixels and the value of P the pixel intensity. Therefore, we have
only partial information about the function P. Not only that, this paltry amount of data is
often polluted with noise.

We can see a connection between the Gaussian blurring and the heat equation here. This
is the solution to the heat equation which can be formulated as follows

∂f ∂ 2 f
=
∂t ∂x 2
f ( x,0) = s( x)
1 ∞  u2 
f ( x, t ) = ∫  − 4t
4πt −∞
exp  ⋅ s( x − u )du

Now if we set t = σ , we have f ( x, σ ) = P( x)


So recovering s is equivalent to solving for f (x,0) in the problem given below:
∂f ∂ 2 f
=
∂t ∂x 2
f ( x, σ ) = P ( x )
However, we will find that if we work on a finite domain and set zero Neumann
conditions, we will still get good results.
∂f ∂ 2 f 
= 2 
∂t ∂x 
∂f 
(0, t ) = 0 
∂x  ……………………………………………….(2)
∂f 
(1, t ) = 0 
∂x 
f ( x,σ ) = P( x)

3. Two Approaches

3.1 Numerical Approach

Although the original problem is posed on the entire real line, as noted above, we will
invert the problem on an interval containing the image of interest by imposing zero
Neumann boundary conditions.

The forward diffusion problem on [0,1] with zero Neumann boundary condition is
represented by the following set of equations.
∂f ∂ 2 f 
= 2 
∂t ∂x 
∂f 
(0, t ) = 0 
∂x  …………………………………………………. (3)
∂f 
(1, t ) = 0 
∂x 
f ( x,0) = s ( x)
We will call the solution operator for this set to be T (t ) .
That is, the solution is f ( x, t ) = T (t ) s ( x) .
In principle we can recover the original image s (x ) by solving the
problem s ( x) = (T (σ )) −1 P( x) .
An approximation for T (σ ) is as follows:
T (t ) s ( x) = ∑ j =0 exp( −π 2 j 2 t ) A j cos( jπx)

≈ ∑ j =0 exp(−π 2 j 2 t ) A j cos( jπx)


K

where the Aj are the Fourier cosine coefficients of s , the original.


Since we know the values of P at finite number of points, we will have
P ( xi ) = ∑ j =0 exp(− j 2π 2σ ) A j cos( jπxi )
K

Plot (1) is the graph of the outputs from Matlab program run on Image1.
As we can see from the plot, we did get a better reconstruction of our original image.

3.2 Regularization using Quasireversibility

Using the method of quasireversibility, we perturb the operator to obtain a well-posed


problem which approximates the original image. We remodeled the problem by changing
the final condition which, now, depends on a small parameter α .
∂f ∂ 2 f
=
∂t ∂x 2
∂f
(0, t ) = 0
∂x
∂f
(1, t ) = 0
∂x
α f ( x,0) + f ( x, σ ) = P( x)

⇒ (α I + T (σ )) f ( x,0) = P( x)

and the discrete problem becomes


K
P ( xi ) = ∑ Aj ⋅ (α + exp(− j 2π 2σ )) ⋅ cos( jπ xi ).
j =0

Plots (2), (3) and (4) are the outputs from Matlab programs run on images 1, 2 and 3. We
see that using quasireversibility we get a better approximation of our images around
particular α and σ values.

3.3 Adding Random Noise

Random noise with zero mean and standard deviation 5% is added to the deblurred image
and the performance of both the techniques discussed above is tested. The results, Plots
(5) and (6), show that even in the presence of noise, the quasireversibility method gave
better results.

3.4 Optimal alpha selection

As a final test, we have written an algorithm to get the optimal α value where we can get
best reconstruction of the blurred image, assuming we have partial information about the
original image. The in-built Matlab function fminsearch is used to find the optimal α
value such the mean-squared error between those pixels of original and deblurred image
is low.
4. Plots
Plot (1) showing original, blurred, and deblurred outputs for Image1

Plot (2) showing outputs for Image1 (with Quasi reversibility)


Plot (3) showing original, blurred, and deblurred outputs for Image2

Plot (4) showing original, blurred, and deblurred outputs for Image3
Plot (5) showing outputs with noise added to Blurred Image1

Plot (6) showing outputs with noise added to Blurred Image1


(using Quasireversibility)
References

[1] Robert A. Hummel, B. Kimia and Steven W. Zucker, Deblurring Gaussian Blur,
Computer Vision, Graphics, and Image Processing 38, 66-80 (1987).

[2] Gordon W. Clark and Seth F. Oppenheimer, Quasireversibility Methods for Non-
Well-Posed Problems, Electronic Journal of Differential Equations.

[3] Stacey Levine, Yunmei Chen and Jon Stanich, Functionals with Nonstandard Growth
in Image Processing,

[4] Dennis G. Zill and Michael R. Cullen, Differential Equations with Boundary-Value
Problems, Fourth Edition.

Anda mungkin juga menyukai