In this problem, given an image we perform the following tasks:
Let us begin by first importing the necessary libraries and defining a few helper functions, before we approach the solutions.
In [2]:
%matplotlib inline
import cv2
import numpy as np
import matplotlib.pyplot as plt
import cmath
The above libraries are necessary for reading image, performing array and matrix operations, visualising the output plots and for handling complex numbers generated by the fourier transform.
Next, we shall define all the necessary functions for solving the given problem.
In [3]:
# Function to perform 2-D Discrete Fourier transform on a given image(matrix) of size M x N.
def two_dim_fourier_transform(img):
M, N = img.shape[:2]
dft_rep = [[0.0 for k in range(M)] for l in range(N)]
for k in range(M):
for l in range(N):
temp_sum = 0.0
for m in range(M):
for n in range(N):
e = cmath.exp(- 1j * 2 * cmath.pi * (float(k * m) / M + float(l * n) / N))
temp_sum += img[m][n] * e
dft_rep[l][k] = temp_sum
return dft_rep
In [4]:
# Function to perform 2-D Inverse Fourier Transform on a given Fourier image matrix.
def two_dim_inv_fourier_transform(fourier):
M = len(fourier)
N = len(fourier[0])
idft_rep = [[0.0 for k in range(M)] for l in range(N)]
for k in range(M):
for l in range(N):
temp_sum = 0.0
for m in range(M):
for n in range(N):
e = cmath.exp(1j * 2 * cmath.pi * (float(k * m) / M + float(l * n) / N))
temp_sum += fourier[m][n] * e
idft_rep[l][k] = temp_sum/(M*N)
return idft_rep
In [5]:
# Function to centre the obtained Fourier transform image. Since Fourier transform is periodic, we can represent
# by relocating the centre of the image for better representation.(Analogous to fftshift)
def img_preprocess_centred_fourier(img):
M, N = img.shape[:2]
processed_img = np.zeros((M,N))
for x in range(M/2):
for y in range(N/2):
processed_img[x][y] = img[M/2 + x][N/2 + y]
for x in range(M/2+1, M):
for y in range(N/2+1, N):
processed_img[x][y] = img[x - M/2 ][y - N/2]
for x in range(M/2+1, M):
for y in range(N/2):
processed_img[x][y] = img[x - M/2][y + N/2]
for x in range(M/2):
for y in range(N/2+1, N/2):
processed_img[x][y] = img[x + M/2][y - N/2]
return processed_img
In [18]:
# If we have an odd sized image, we convert it into an even sized image.
def preprocess_odd_images(img):
M, N = img.shape[:2]
if M % 2 == 1 and N % 2 == 1:
return img[1:][1:]
elif M%2 == 1:
return img[1:][:]
elif N%2 == 1:
return img[:][1:]
else:
return img
In [6]:
# Function to plot a labeled image
def plot_input(img, title):
plt.imshow(img, cmap = 'gray')
plt.title(title), plt.xticks([]), plt.yticks([])
plt.show()
In [7]:
# Function to compute the Mean Squared Error between two input matrices.
def min_sqr_err(mat1, mat2):
min_sq = 0
M, N = mat1.shape[:2]
for i in range(M):
for j in range(N):
min_sq += np.square(mat1[i][j] - mat2[i][j])
return min_sq
Let us begin the solution by loading an input image. Since the computation of Fourier transform is extremely memory intensive, we choose a very small image (40 x 40 pixels) as input testing image.
In [19]:
img = cv2.imread('./fb_test.jpg', 0)
img = preprocess_odd_images(img)
plot_input(img,'Input image')
In [9]:
# Details of the input image
M, N = img.shape[:2]
print M,N
In [20]:
processed_img = img_preprocess_centred_fourier(img)
dft_rep = two_dim_fourier_transform(img)
centred_dft_rep = two_dim_fourier_transform(processed_img)
# Visualizing the Fourier Transform in Logarithmic Scale
plot_input(10*np.log(1+np.abs(centred_dft_rep)), 'Fourier Transform')
As we see, above is the fourier transform of the input image.
Now let us try and see if we can reconstruct the original image, by performing Inverse fourier transform on the above obtained image in the Fourier domain.
In [16]:
idft_rep = two_dim_inv_fourier_transform(dft_rep)
plot_input(np.abs(idft_rep), 'Inverse Fourier Image')
As we see, the image obtained by performing the inverse fourier transform of the obtained fourier image, is very similar to the input image.
But let us compute the Mean Squared Error(MSE) between the original image and the reconstructed image, to verify that the images are identical or not.
In [17]:
min_sq_error = min_sqr_err(img, np.abs(idft_rep))
print "Mean squared Error between the reconstructed and original image: ", min_sq_error
The error is extremely negligible, of the magnitude of 10^(-21) ~ Zero.