1 minute read

이미지나 신호를 주파수 도메인에서 처리하는 기법

저주파 또는 고주파 성분을 강조하거나 억제함으로써 특정 특징을 추출하거나 노이즈를 제거. 저역 통과 필터(Low-pass filter)는 이미지의 평활화에, 고역 통과 필터(High-pass filter)는 에지 강조에 주로 사용

Frequency Domain


Frequency Domain Filter

Frequency Domain Filter

Filtering fundamentals

  • $g(x,y)=\textrm{Real} [\Im^{-1}[H(\mu ,\nu )F(\mu ,\nu)]]$
    • $\Im^{-1}$: IDFT
    • $F(\mu ,\nu)$: DFT of the input image $f(x,y)$
    • $H(\mu ,\nu)$: Filter transfer function
    • $g(x,y)$: Filtered (output) image
    • $H(\mu ,\nu )F(\mu ,\nu)$: Elementwise multiplication
  • 저 주파수: 이미지에서 천천히 변하는 intensity 구성 요소
  • 고 주파수: Intensity 의 급격한 변화
    • e.g., edge, noise
  • $H(\mu, \nu)$
    • 저 주파수 통과 및 고 주파수 감소: Blur
    • 고 주파수 통과 및 저 주파수 감소: Sharpening (선명한 세부 정보 강화)
  • image


Frequency Domain Filtering Steps

  1. Given an input image $f(x,y)$ of size M x N, obtain padding sizes $P = 2M$ and $Q = 2N$
  2. Form a padding image $f_{p}(x,y)$ of size P x Q (zero, mirror, or replicate padding)
  3. Multiply $f_{p}(x,y)$ by $(-1)^{x+y}$ to center the FT
  4. Compute DFT $F(\mu ,\nu)$ of an image from Step 3
  5. Construct a real, symmetric filter transfer function $H(\mu, \nu)$ of size P x Q with center at (P/2 , Q/2)
  6. Elementwise multiplication: $G(\mu, \nu) = H(\mu, \nu)F(\mu, \nu)$
  7. Obtain the filtered image of size P x Q by computing the IDFT of $G(\mu, \nu)g_{p}(x,y) = (\textrm{Real} [\Im^{-1}[H(\mu ,\nu )F(\mu ,\nu)]])(-1)^{x+y}$
  8. Obtain the final filtered result $g(x,y)$ of size M x N by extracting the M x N regions from the top, left quadrant of $g_{p}(x,y)$

image


Smoothing using low-pass

  • Attenuate high frequency via low-pass filtering
  • Ideal LPF: cut off all frequencies outside a circle of radius
  • $H(\mu ,\nu) = 1 : \textrm{ if } D(\mu ,\nu) \leq D_0 \ , 0 : \textrm{ if } D(\mu ,\nu) > D_0 $
    • $D_{0}$: Cut-off Frequency
    • $D(u,v)$: (u,v) 지점 사이의 거리 in the frequency domain and the center

image

image


Ideal LPF

image image


Gaussian LPF

image


Butterworth LPF

image

image image


Sharpening using high-pass

Leave a comment