Unit 2: Image Enhancement in
Frequency Domain; Image Restoration
Fourier Transform and Frequency Domain
The Fourier Transform (FT) converts spatial domain images into the frequency domain. In
this domain, an image is represented as a sum of sinusoids of different frequencies. The FT
is essential in image analysis and filtering, especially for periodic patterns.
Filtering in Frequency Domain
Filtering involves modifying the frequency components of an image. Low-pass filters (LPF)
retain low-frequency components, smoothing the image, while high-pass filters (HPF)
enhance edges and fine details. These filters are applied by multiplying the image FT with a
filter mask.
Spatial and Frequency Domain Correspondence
Filtering in the spatial domain using convolution is equivalent to multiplication in the
frequency domain due to the convolution theorem. This provides computational advantages
for large filters or large images.
Gaussian Filters
Gaussian Low-pass filters (GLPF) reduce high-frequency noise while preserving edges
better than ideal filters. Gaussian High-pass filters (GHPF) suppress low-frequency
components, enhancing fine image details. Their smooth roll-off prevents artifacts.
Homomorphic Filtering
This technique separates illumination and reflectance in an image, enhancing contrast and
brightness. It involves logarithmic transformation, filtering in the frequency domain, and
inverse transformation.
Image Restoration - Noise Models
Image restoration aims to recover the original image from a degraded version. Common
noise models include Gaussian, salt-and-pepper, and Poisson noise. Modeling noise helps
choose appropriate restoration techniques.
Spatial Filtering for Noise Removal
Mean filters (arithmetic and geometric) reduce random noise by averaging neighboring
pixels. Order-statistic filters like median, max, and min filters are effective for impulse noise.
Periodic Noise Reduction in Frequency Domain
Periodic noise appears as distinct peaks in the frequency domain. Band-reject or band-pass
filters can selectively suppress or retain frequency bands to reduce such noise.
Minimum Mean-Square Error (MMSE) Restoration
MMSE minimizes the average squared error between the restored and original image. It
requires knowledge of noise and original image statistics. It's effective for Gaussian noise
but computationally intensive.