Nadav Borenstein - 1 year ago 147

Python Question

I am trying to implement in python some functions that transform images to their fourier domain and vice-versa for image processing tasks.

I implemented the 2D-DFT using repeated 1D-DFT, and it worked fine, but when I tried to Implement 2D inverse DFT using repeated inverse 1D-DFT, some weird problem occurred - When I transform an image to its fourier domain and then back to the image domain, it looks like the image was reflected and merged with its reflection, as can be seen here:

this is the input:

and this is the output

This is the function that is responsible for the mess:

`def IDFT2(fourier_image):`

image = np.zeros(fourier_image.shape)

for col in range(image.shape[1]):

image[:, col] = IDFT1(fourier_image[:, col])

for row in range(image.shape[0]):

image[row, :] = IDFT1(image[row,:])

return image

What did I do wrong? I am pretty sure that IDFT1 works fine, and so is the regular 2D-DFT.

the code for the IDFT is:

`def create_DFT_matrix(N):`

w = np.exp(-(1/N) * 2 * np.pi * 1j)

DFT_matrix = np.array([range(N)] * N)

DFT_matrix = (DFT_matrix.T.dot(DFT_matrix) / N).astype(np.int64)

DFT_matrix = w ** DFT_matrix

return DFT_matrix

def IDFT(fourier_signal):

DFT_matrix = create_DFT_matrix(fourier_signal.shape[0])

IDFT_matrix = np.linalg.inv(DFT_matrix)

signal = matrix.dot(fourier_signal.T)

return np.real(signal).astype(np.float32)

Recommended for you: Get network issues from **WhatsUp Gold**. **Not end users.**

Answer Source

I do not use **Python** so I am not confident to analyze your code but my bet is that you most likely forget to implement complex values at some stage....

it should be:

**DFT**rows from real to complex domain**DFT**columns of result from complex to complex domain- apply normalization if needed
- any or none processing ...
**iDFT**rows from complex to complex domain**iDFT**columns of result from complex to real domain- apply normalization if needed

if you use just real to complex domain **DFT/iDFT** in the second passes (bullets **#2,#6**) then it would create the mirroring because **DFT** of real values is a mirrored sequence ... Btw. it does not matter if you process rows or columns first ... also you can process rows first in **DFT** and columns first in **iDFT** the result should be the same +/- floating errors ...

for more info see

and all the sub-links there especially the `2D FFT and wrapping example`

so you can compare your results with something working

Recommended from our users: **Dynamic Network Monitoring from WhatsUp Gold from IPSwitch**. ** Free Download**