Image filtering in frequency domain python

asked 2015-05-21 03:06:16 -0600

fery gravatar image

updated 2015-05-21 03:55:43 -0600

Hi everybody,

I am new in programming and I would like to apply a filter on an image in frequency domain. actually, its from a paper and i want to re implement it. its the formula: im_out= (1+ 5*((1-e^-f)/f)) * im_in and here are my codes:

im = cv2.imread('lena.jpg',0)

x=im.shape[0]

y=im.shape[1]

fft= np.fft.fft2(im)

fshift= np.fft.fftshift(fft)

freqx = np.fft.fftfreq(fft.shape[0])

freqy = np.fft.fftfreq(fft.shape[1])

zr = freqx==0

freqx[zr]=0.000000001

for i in xrange(x):

  for j in xrange(y):

    filter = 1+(5*(1-np.e**(-1*(np.sqrt((freqx)**2 + (freqy)**2))))/(np.sqrt((freqx)**2 + (freqy)**2)))

fim= fshift* filter

fishift= np.fft.ifftshift(fim)

imback=np.fft.ifft2(fishift)

imback=np.uint8(np.real(imback))

plt.subplot(221)

plt.imshow(im,cmap='gray')

plt.subplot(222)

plt.imshow(imback,cmap='gray')

plt.show()

but i get the image without any visible changes, it should be kind of low pass filter. now I am wondering if its correct to use np.fft.fftfreq to find "spatial frequency in the image plane". or I should use distance from center as "f"?!!! in the paper they said "f" is spatial frequency of the image plane!!!!

could anybody help me plz !!!

edit retag flag offensive close merge delete

Comments

I don't understand python code but you don't need fftshift. Normalize frequency fc for an image of nbCols and nbRows is given by fcx=i/nbCols and fcy = j/nbRows your f value is something like f=sqrt(fcx^2+fcy^2) Becarefull for real signal you have a hermtian symetry for your filter

LBerger gravatar imageLBerger ( 2015-05-21 05:06:35 -0600 )edit

LBerger, thx for your answer. u are telling that i don't need to shift the image to have DC in the center. and just divide the pixel number to the number of rows and col ? what do u mean by hermtian symetry? how it affects?

fery gravatar imagefery ( 2015-05-21 05:53:34 -0600 )edit

It was Hermitianif i(x,y) is your image with real number (not complex)and I=F(i(x,y))is fourier transform of i then I(x,y)=conjugate(I(nbCols-x,nbrows-y)) You can find this in signal processing book. Your filter must have this symetry Fftshift is used for graphic representation : we like to have O(0,0) at the center of the figure

LBerger gravatar imageLBerger ( 2015-05-21 06:44:16 -0600 )edit

Ah got it, its symmetric. but its not working like this i won't get a blur image it just make the image a little darker just it!!!!

fery gravatar imagefery ( 2015-05-21 07:23:46 -0600 )edit

Yes Frequency response of your filter

LBerger gravatar imageLBerger ( 2015-05-21 07:40:21 -0600 )edit

you mean that the filter is not for blurring?????? it should blur the image, but just make it a little darker!!!

fery gravatar imagefery ( 2015-05-21 08:29:07 -0600 )edit