While I was reading SIFT paper(Lowe, 2004), I came across the method that he apply "heat diffusion equation" to Gaussian function to derive that $$ \frac{∂G}{∂σ} = σ∇^2G $$ I searched Wikipedia and found that original heat equation is

$$ \frac{∂u}{∂t} = σ∇^2u $$ I wonder how can time derivative equation be applied to sigma derivative equation. Is there anybody to understand me?

  • $\begingroup$ I don't understand what you're after. LoG usually used in Image Processing to apply a filter to approximate smooth Laplacian of the Image. $\endgroup$ – Royi Jul 19 '18 at 6:43
  • $\begingroup$ Lowe approximated LoG to DoG by using "heat diffusion equation". I don't understand why u function(in the heat equation) can be substituted with Gaussian function. $\endgroup$ – uninopkn Jul 19 '18 at 7:05

I'm not sure I fully understood what's the issue you're having.
Yet I will show a simple property of the Gaussian filter which might make things clearer.

For simplicity, I will use 1D Signal.
Yet it is easy to extend it into 2D.

The Problem

For $ u \in \mathbb{R} $ (1D, unbounded domain), show that a Gaussian convolution with the initial condition solves the linear diffusion equation:

$$ {u}_{t} = {u}_{xx}, \quad u \left( t = 0; x \right) = f \left( x \right) $$

That is:

$$ u \left( t; x \right) = f \left( x \right) \ast {g}_{ \sigma \left( t \right)} \left( x \right) $$

Where $ \ast $ denotes convolution.

Find the function $ F \left( \cdot \right) $ relating the evolution time $ t $ to the Gaussian standard deviation $ \sigma \: : \: \sigma = F \left( t \right) $.


Looking at:

$$ \frac{\partial u \left( x, t \right ) }{\partial t } = \frac{\partial^2 u \left( x, t \right ) }{\partial x^2} , \; u \left( x, t = 0 \right) = f \left( x \right) $$

Defining $ U \left( s, t \right) $ as the Fourier Transform of $ u \left( x, t \right) $ with respect to the parameter $ x $.
This leads to:

$$ \begin{align} \mathfrak{F} \left( \frac{\partial u }{\partial t} \right) & = \frac{\partial U \left( s, t \right ) }{\partial t} \\ \mathfrak{F} \left( \frac{\partial^2 u}{\partial x^2} \right) & = -4 {\pi}^{2} {s}^{2} U \left( s, t \right ) \\ \mathfrak{F} \left( f \left( x \right) \right) & = F \left( s \right) \end{align} $$

Hence the equation in the Fourier domain is given by:

$$ \frac{\partial U \left( s, t \right ) }{\partial t} = -4 {\pi}^{2} {s}^{2} U \left( s, t \right ) , \; U \left( s, t = 0 \right) = F \left( s \right) $$

The solution of this Linear Ordinary Differential Equation is given by:

$$ U \left( s, t \right) = F \left( s \right) {e}^{-4 {\pi}^{2} {S}^{2} t } $$

Hence the solution in the time domain is given by (Using the identity $ \mathfrak{F} \left( {e}^{- \alpha {x}^{2}} \right) = \sqrt{\frac{\pi}{\alpha}} {e}^{- \frac{{\left( \pi f \right)}^{2}}{\alpha}} $):

$$ u \left( x, t \right) = \frac{1}{\sqrt{2 \pi 2t}} {e}^{- \frac{{x}^{2}}{4 t}} = \frac{1}{\sqrt{2 \pi {\sigma \left( t \right)}^{2}}} {e}^{- \frac{{x}^{2}}{2 {\sigma \left( t \right)}^{2}}} $$

Comparing it to the classic Gaussian Function yields that $ \sigma \left( t \right) = \sqrt{\left( 2 t \right)} $.

This is a solution of mine to a problem I was given in a course named Variational Methods in Image Processing.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.