1. 程式人生 > >Gabor函式和Gabor濾波器的原理和實現

Gabor函式和Gabor濾波器的原理和實現

Gabor函式
  Gabor變換屬於加窗傅立葉變換,Gabor函式可以在頻域不同尺度、不同方向上提取相關的特徵。另外Gabor函式與人眼的生物作用相仿,所以經常用作紋理識別上,並取得了較好的效果。二維Gabor函式可以表示為:
  其中:
  v的取值決定了Gabor濾波的波長,u的取值表示Gabor核函式的方向,K表示總的方向數。引數決定了高斯視窗的大小,這裡取。程式中取4個頻率(v=0, 1, ..., 3),8個方向(即K=8,u=0, 1, ... ,7),共32個Gabor核函式。不同頻率不同方向的Gabor函式可通過下圖表示:
  圖片來源:GaborFilter.html
  圖片來源:http://www.bmva.ac.uk/bmvc/1997/papers/033/node2.html
  三、程式碼實現
  Gabor函式是復值函式,因此在運算過程中要分別計算其實部和虛部。程式碼如下:

  private void CalculateKernel(int Orientation, int Frequency)
  {
  double real, img;
  for(int x = -(GaborWidth-1)/2; x<(GaborWidth-1)/2+1; x++)
  for(int y = -(GaborHeight-1)/2; y<(GaborHeight-1)/2+1; y++)
  {
  real = KernelRealPart(x, y, Orientation, Frequency);
  img = KernelImgPart(x, y, Orientation, Frequency);
  KernelFFT2[(x+(GaborWidth-1)/2) + 256 * (y+(GaborHeight-1)/2)].Re = real;
  KernelFFT2[(x+(GaborWidth-1)/2) + 256 * (y+(GaborHeight-1)/2)].Im = img;
  }
  }
  private double KernelRealPart(int x, int y, int Orientation, int Frequency)
  {
  double U, V;
  double Sigma, Kv, Qu;
  double tmp1, tmp2;
  U = Orientation;
  V = Frequency;
  Sigma = 2 * Math.PI * Math.PI;
  Kv = Math.PI * Math.Exp((-(V+2)/2)*Math.Log(2, Math.E));
  Qu = U * Math.PI / 8;
  tmp1 = Math.Exp(-(Kv * Kv * ( x*x + y*y)/(2 * Sigma)));
  tmp2 = Math.Cos(Kv * Math.Cos(Qu) * x + Kv * Math.Sin(Qu) * y) - Math.Exp(-(Sigma/2));
  return tmp1 * tmp2 * Kv * Kv / Sigma;
  }
  private double KernelImgPart(int x, int y, int Orientation, int Frequency)
  {
  double U, V;
  double Sigma, Kv, Qu;
  double tmp1, tmp2;
  U = Orientation;
  V = Frequency;
  Sigma = 2 * Math.PI * Math.PI;
  Kv = Math.PI * Math.Exp((-(V+2)/2)*Math.Log(2, Math.E));
  Qu = U * Math.PI / 8;
  tmp1 = Math.Exp(-(Kv * Kv * ( x*x + y*y)/(2 * Sigma)));
  tmp2 = Math.Sin(Kv * Math.Cos(Qu) * x + Kv * Math.Sin(Qu) * y) - Math.Exp(-(Sigma/2));
  return tmp1 * tmp2 * Kv * Kv / Sigma;
  }
有了Gabor核函式後就可以採用前文中提到的“離散二維疊加和卷積”或“快速傅立葉變換卷積”的方法求解Gabor變換,並對變換結果求均值和方差作為提取的特徵。32個Gabor核函式對應32次變換可以提取64個特徵(包括均值和方差)。由於整個變換過程程式碼比較複雜,這裡僅提供測試程式碼供下載。該程式碼僅計算了一個101×101尺寸的Gabor函式變換,得到均值和方差。程式碼採用兩種卷積計算方式,從結果中可以看出,快速傅立葉變換卷積的效率是離散二維疊加和卷積的近50倍。
-----------------------------------------------------------------------

最近忙著論文,需要Gabor濾波程式碼,可是網上總找不到合適的程式碼,於是就自己編了一個,不當之處請指點。參考論文為 L. Wiskott,J. M. Fellous,N. Kruger,C. v. d.Malsburg. Face Recognition by Elastic Bunch Graph Matching,IEEE Trans. On PAMI,Vol.19,No.7,pp775-779,1997

首先實現濾波器:

function [bank] = do_createfilterbank(imsize,varargin)

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% 函式實現:建立Gabor 濾波組
%
%%% 必選引數:
%   imsize - 影象大小
%%% 可選引數:
%   freqnum — 頻率數目
%   orientnum — 方向數目
%   f       —   頻率域中的取樣步長
%   kmax    —   最大的取樣頻率
%   sigma   —   高斯窗的寬度與波向量長度的比率
%
%%% 返回結果:
%   bank
%           .freq        —     濾波頻率
%           .orient      —     濾波方向
%           .filter      —     Gabor濾波
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

conf = struct(...,
    'freqnum',3,...
    'orientnum',6,...
    'f',sqrt(2),...
    'kmax',(pi/2),...
    'sigma',(sqrt(2)*pi) ...
    );

conf = do_getargm(conf,varargin);

bank = cell(1,conf.freqnum*conf.orientnum);
for f0=1:conf.freqnum
    fprintf('處理頻率 %d \n', f0);
    for o0=1:conf.orientnum
        [filter_,freq_,orient_] = do_gabor(imsize,(f0-1),(o0-1),conf.kmax,conf.f,conf.sigma,conf.orientnum);
        bank{(f0-1)*conf.orientnum + o0}.freq = freq_; %以orient增序排列
        bank{(f0-1)*conf.orientnum + o0}.filter = filter_;
        bank{(f0-1)*conf.orientnum + o0}.orient = orient_;
    end
end

for ind = 1:length(bank)
    bank{ind}.filter=fftshift(bank{ind}.filter);
end

function [filter,Kv,Phiu] = do_gabor (imsize,nu,mu,Kmax,f,sigma,orientnum)

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% 函式實現: 建立Gabor濾波
%
%%% 引數:
%   imsize     : 濾波的大小(即影象大小)
%   nu         : 頻率編號 [0 ...freqnum-1];
%   mu         : 方向編號 [0...orientnum-1]
%   Kmax       : 最大的取樣頻率
%   f          : 頻率域中的取樣步長
%   sigma      : 高斯窗的寬度與波向量長度的比率
%   orientnum : 方向總數
%
%%% 返回值:
%   filter :    濾波
%   Kv    :    頻率大小
%   Phiu :    方向大小
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

rows = imsize(1);
cols = imsize(2);
minrow = fix(-rows/2);
mincol = fix(-cols/2);
row = minrow + (0:rows-1);
col = mincol + (0:cols-1);
[X,Y] = meshgrid(col,row);

Kv = Kmax/f^nu;
Phiu = pi * mu /orientnum;
K = Kv * exp(i * Phiu);

F1 = (Kv ^ 2)/ (sigma^2) * exp(-Kv^2 * abs(X.^2 + Y.^2) / (2*sigma^2)) ;
F2 = exp(i * (real(K) * X + imag(K) * Y)) - exp(-sigma^2/2);
filter = F1.* F2;

Gabor 濾波實現(1)已經建立了Gabor濾波組,現在可以使用該濾波組對影象進行轉換,得到振幅和相位。

function [result] = do_filterwithbank(im,bank)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% 函式實現:對影象使用Gabor濾波組進行轉換換
%
%%% 引數:
%   im       — 被轉換的影象
%   bank        — 由函式do_createfilterbank得到的濾波組
%
%%% 返回:
%   result                     — 影象被轉換後的結果
%       .amplitudes            — 不同畫素點的振幅向量
%       .phases                — 不同畫素點的相位向量
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

[N1 N2] = size(im);
N3 = length(bank);
phases = zeros(N1,N2,N3);
amplitudes = zeros(N1,N2,N3);
imagefft = fft2(im);

for ind = 1:N3
    fprintf('正在處理濾波 %d \n',ind);
     temp = ifft2(imagefft .* bank{ind}.filter);
     phases(:,:,ind) = angle(temp);
     amplitudes(:,:,ind) = abs(temp);
end
result.phases = phases;
result.amplitudes = amplitudes;
整個程式可以如下使用。 im = imread('image.jpg'); im = rgb2gray(im); bank = do_createfilterbank(size(im)); result = do_filterwithbank(im,bank);