I have an elementary function $p:\mathbb{R}^2\rightarrow\mathbb{R}$ which (locally) represents an image. It's a polynomial, and its the result of the following 2D convolution:

$$p=f\star G\star \mathbf{1}_{A}$$

where $G(x,y)$ is a 2D Gaussian, and $\mathbf{1}_A$ 'box' function:

$$G(x,y)=e^{-\frac{(x^2+y^2)}{2\sigma^2}}, \quad\mathbf{1}_A(x,y)=1 \text{ for }x \text{ and } y < a, 0\text{ otherwise.}$$

Is there a closed form for a function with which I can convolve $p$, which results in $f$?

Currently I am 'avoiding' the use of such a filter by performing Fourier deconvolution, but mathematically this opens a huge can of worms.


The FT of the convolutional inverse is the point-wise inverse in the FD. So the convolutional inverse of a gaussian might have numerical problems. The FT of a gaussian is a gaussian. Its point-wise inverse thus can be large in many places.

For a box function, the situation doesn't look all that good either. Its FT is a variant on $\operatorname{sinc}$, which has many zeroes and loci of low magnitude.

You can certainly determine the convolutional inverse of a gaussian, but the inverse of a box function is not even a function. Convolving with a box will cancel any function that is periodic with a period multiple of the edge length.


The problem with deconvolution is that you amplify quantization noise and other stuff and the inverse might not even have a stable representation. Look up "Wiener filtering": it is least-square error inverse filtering with a signal and noise model and can usually be done easily enough in the frequency domain.


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.