Linear Diffusion Flows¶


This tours studies linear diffusion PDEs, a.k.a. the heat equation. A good reference for diffusion flows in image processing is Weickert98.

In [2]:
addpath('toolbox_signal')


Heat Diffusion¶

The heat equation reads $$\forall t>0, \quad \pd{f_t}{t} = \nabla f_t$$ for a function $f_t : \RR^2 \rightarrow \RR$ and where $f_0$ (the solution at initial time $t=0$) is given.

The Laplacian operator reads $$\Delta f = \pdd{f}{x_1} + \pdd{f}{x_2}.$$

The flow is discretized in space by considering a discrete image of $N = n \times n$ pixels.

In [3]:
n = 256;


Load an image $f_0 \in \RR^N$, that will be used to initialize the flow at time $t=0$.

In [4]:
name = 'hibiscus';
f0 = rescale( sum(f0,3) );


Display it.

In [5]:
clf;
imageplot(f0);


The flow is discretized in time using an explicit time-stepping $$f^{(\ell+1)} = f^{(\ell)} + \tau \Delta f^{(\ell)}.$$ We use finite difference Laplacian $$(\Delta f)_i = \frac{1}{h^2}\pa{ f_{i_1+1,i_2}+f_{i_1-1,i_2}+f_{i_1,i_2+1}+f_{i_1,i_2-1}-4f_j }$$ where we assume periodic boundary conditions, and where $h = 1/N$ is the spacial step size.

In [6]:
h = 1/n;


The step size $\tau$ should satisfy $$\tau < \frac{h^2}{4}$$ for the discretized flow to be stable.

The discrete solution $f^{(\ell)}$ converges to the continuous solution $f_t$ at time $t = \tau \ell$ if both $\tau \rightarrow 0$ and $h \rightarrow 0$ under the condition $\tau/h^2 < 1/4$.

Select a small enough step size.

In [7]:
tau = .5 * h^2/4;


Final time.

In [8]:
T = 1e-3;


Number of iterations.

In [9]:
niter = ceil(T/tau);


Initialize the diffusion at time $t=0$.

In [10]:
f = f0;


One step of discrete diffusion.

In [11]:
f = f + tau * delta(f);


Exercise 1

Compute the solution to the heat equation.

In [12]:
exo1()

In [13]:
%% Insert your code here.


Explicit Solution using Convolution¶

The solution to the heat equation can be computed using a convolution $$\forall t>0, \quad f_t = f_0 \star h_t$$ where $\star$ denotes the convolution of continuous functions $$f \star h(x) = \int_{\RR^2} f(y) g(x-y) d y$$ and $h_t$ is a Gaussian kernel of width $\sqrt{t}$ $$h_t(x) = \frac{1}{4 \pi t} e^{ -\frac{\norm{x}^2}{4t} }$$

One can thus approximate the solution using a discrete convolution. Convolutions can be computed in $O(N\log(N))$ operations using the FFT, since $$g = f \star h \qarrq \forall \om, \quad \hat g(\om) = \hat f(\om) \hat h(\om).$$

In [14]:
cconv = @(f,h)real(ifft2(fft2(f).*fft2(h)));


Define a discrete Gaussian blurring kernel of width $\sqrt{t}$.

In [15]:
t = [0:n/2 -n/2+1:-1];
[X2,X1] = meshgrid(t,t);
normalize = @(h)h/sum(h(:));
h = @(t)normalize( exp( -(X1.^2+X2.^2)/(4*t) ) );


Define blurring operator.

In [16]:
heat = @(f, t)cconv(f,h(t));


Example of blurring.

In [17]:
clf;
imageplot(heat(f0,2));


Exercise 2

Display the heat convolution for increasing values of $t$.

In [18]:
exo2()

In [19]:
%% Insert your code here.