MATLAB Feature Extraction experiment process

First, standardize the images in the face database and cut out the face area based on the two eyes, as shown in:

The MATLAB program is as follows:

% Function Description: The face size normalization manual click to automatically cut the image as the specified size function unitery (height, width, m) % wdith is the width of the cut image. Height is the length of the cut image. M indicates the number of images. Clear; close all; Height = 490; % set image size width = 640; M = 12; % set numbers of image % set path of image mydir = 'H: \ CMU emojis library \ Cohn-kanade \ s010 \ 001 \ '; % set suffix dirs = Dir ([mydir ,'*. PNG ']); for I = 1: m STR = [mydir, dirs (I ). name]; EVAL ('oriimg = imread (STR); '); % execute the string to read IMG % into the image figure, imshow (oriimg), [x, y] = ginput (2); % manually obtain the human eye E1, E2 coordinate d = x ()-X ); % calculate the width between two eyes. d ox = sum (X)/2; Oy = sum (y)/2; % calculate the center O (OX, Oy) of E1 and E2) i1 = imcdrop (oriimg, [ox-0.9 * D Oy-0.5 * D 1.8 * D 2 * D]); % cut face STR = strcat ('H: \ CMU emoticons library \ Cohn-kanade \ s010 \ 001 \ Standard \ ',int2str( I }'.bmp '); EVAL ('imwrite (I1, STR );'); % execute a string to read IMG % into image % to save the normalized Face Image close all; End

 

After standardization, the task is to partition up and down, and cut off evenly.

Shape:

The MATLAB program is as follows:

function ShangXiaFenKuai    mydir = 'C:\Users\user\Desktop\Database\CMU\test_test\';    DIRS = dir([mydir,'*bmp']);    for i=1:18    if ~DIRS(i).isdir        filename = [mydir,DIRS(i).name];        disp(filename)        I = imread(filename);        ImageUp = I(1:50,:);        ImageDown = I(51:100,:);        savefileUp = [mydir,'Up\',DIRS(i).name];        imwrite(ImageUp,savefileUp);        savefileDown = [mydir,'Down\',DIRS(i).name];        imwrite(ImageDown,savefileDown);    end    i = i + 1;    endend

Use functions in MATLAB to extract Gabor Wavelet features

Two M files are involved here.

Main. m

Mydir = 'C: \ Users \ User \ Desktop \ data \ '; for I = for J = dir = running mydir,num2str( I #,'-', num2str(j#,'.bmp']; Im = imread (DIR ); % or = rgb2gray (IM); Gim = im2double (IM); disp (DIR); [EIM, OIM, aim] = spatialgabor (Gim, 4.8, 150, 0.4, 0.2, 1); SaveFile = [mydir, 'gabor \ ', num2str(ifolder, '-', num2str(jfolder, '.bmp']; imwrite (aim, SaveFile ); test of endend % Gabor parameter % % Ori = imread ('d: \*. JPG '); OR = rgb2gray (ORI) Gim = im2single (OR); for I = for J = 30: 20: 150 for X = 0.1: 0.1: 1 For Y = 0.1: 0.1: 1 [EIM, OIM, aim] = spatialgabor (Gim, I, j, X, Y, 1); SaveFile = ['d: \ Ba \ alipay'] imwrite (aim, SaveFile ); end end % %

Spatialgabor. m

% SPATIALGABOR - applies single oriented gabor filter to an image%% Usage:%  [Eim, Oim, Aim] =  spatialgabor(im, wavelength, angle, kx, ky, showfilter)%% Arguments:%         im         - Image to be processed.%         wavelength - Wavelength in pixels of Gabor filter to construct%         angle      - Angle of filter in degrees.  An angle of 0 gives a%                      filter that responds to vertical features.%         kx, ky     - Scale factors specifying the filter sigma relative%                      to the wavelength of the filter.  This is done so%                      that the shapes of the filters are invariant to the%                      scale.  kx controls the sigma in the x direction%                      which is along the filter, and hence controls the%                      bandwidth of the filter.  ky controls the sigma%                      across the filter and hence controls the%                      orientational selectivity of the filter. A value of%                      0.5 for both kx and ky is a good starting point.%         showfilter - An optional flag 0/1.  When set an image of the%                      even filter is displayed for inspection.% % Returns:%         Eim - Result from filtering with the even (cosine) Gabor filter%         Oim - Result from filtering with the odd (sine) Gabor filter%         Aim - Amplitude image = sqrt(Eim.^2 + Oim.^2)%% Peter Kovesi  % School of Computer Science & Software Engineering% The University of Western Australia% pk at csse uwa edu au% http://www.csse.uwa.edu.au/~pk%% October 2006function [Eim, Oim, Aim] = spatialgabor(im, wavelength, angle,kx,ky, showfilter)    if nargin == 5        showfilter = 0;    end    im = double(im);    [rows, cols] = size(im);    newim = zeros(rows,cols);    % Construct even and odd Gabor filters    sigmax = wavelength*kx;    sigmay = wavelength*ky;        sze = round(3*max(sigmax,sigmay));    [x,y] = meshgrid(-sze:sze);    evenFilter = exp(-(x.^2/sigmax^2 + y.^2/sigmay^2)/2)...    .*cos(2*pi*(1/wavelength)*x);        oddFilter = exp(-(x.^2/sigmax^2 + y.^2/sigmay^2)/2)...    .*sin(2*pi*(1/wavelength)*x);        evenFilter = imrotate(evenFilter, angle, 'bilinear');    oddFilter = imrotate(oddFilter, angle, 'bilinear');        % Do the filtering    Eim = filter2(evenFilter,im); % Even filter result    Oim = filter2(oddFilter,im);  % Odd filter result    Aim = sqrt(Eim.^2 + Oim.^2);  % Amplitude     %if showfilter % Display filter for inspection        %figure(1), imshow(evenFilter,[]); title('filter');     end    

After the Gabor feature map is obtained, the dimension is reduced. Here, PCA and DCT are used for dimensionality reduction.

PCA. m

Function exepca1 () CLC; % clear screen % % 18 A large matrix % % % dirtest = 'd: \ train \ test-6-3 \ Gabor \ '; dirs_test = Dir ([dirtest ,'*. BMP ']); for I = If ~ Dirs_test (I ). isdir % to be loaded and processed. Here you can get string_test = [dirtest, dirs_test (I ). name]; disp (string_test); % read image image_src_test = imread (string_test); % The following zscore function is processed by row, think of the row vector as a sample % convert the image matrix into a column vector % first transpose the image, then expand by column, and then get the row vector in the transpose !! I1 = image_src_test '; vector_test = I1 (:); image_vector_test = vector_test'; % image_vector = image_src (:); % form a new large matrix if I = 1 image_matrix_test = image_vector_test; else image_matrix_test = [image_matrix_test; image_vector_test]; end image_double_test = im2single (image_matrix_test ); % % merge a large matrix % % % Mydir = 'd: \ train-6-5 \ Gabor-90 \ 'dirs = Dir ([mydir ,'*. BMP ']); for I = If ~ Dirs (I ). isdir % is to be loaded and processed. Here you can get string = [mydir, dirs (I ). name]; disp (string); % read image image_src = imread (string); % The following zscore function is processed by row, think of the row vector as a sample % convert the image matrix into a column vector % first transpose the image, then expand by column, and then get the row vector in the transpose !! I1 = image_src '; vector = I1 (:); image_vector = vector'; % image_vector = image_src (:); % form a new large matrix if I = 1 image_matrix = image_vector; flag = 2; else image_matrix = [image_matrix; image_vector]; end image_double = im2single (image_matrix ); % % xz = zscore (image_double); % evaluate the 'feature value ', 'difference', 'contribution rate ', and 'accumulation contribution rate'. The base of the main requirement is a set of feature values [coeff, score, late NT] = princomp (xz); % [coeff, latent] = pcacov (COV (xz); explained = 100 * latent/sum (latent); [m, n] = size (image_double); Result = cell (n + 1, 4); Result (1, :) = {'feature value ', 'difference', 'contribution rate ', 'accumulative contribution rate '}; Result (2: end, 1) = num2cell (latent); Result (2: End-1, 2) = num2cell (-diff (latent )); result (2: end, 3: 4) = num2cell ([explained, cumsum (explained)]); % % This sentence is restored !!!! Pca_train = image_double * coeff; % the first 100 elements after restoration are taken as the representative !!! Pca400_train = pca_train (:,); filename = require mydir,'imagepca_train.txt ']; disp (filename); dlmwrite (filename, pca400_train ); % % pca_test = image_double_test * coeff; pca400_test = pca_test (:,); filename_test = nvidirtest,'imagepca_test.txt ']; disp (filename_test); dlmwrite (filename_test, pca400_test, '',); End

 

DCT. m

 

% Evaluate the result of DCT transformation retain the first 300 dimension function dctclc; flag = 1; filename = 'C: \ Users \ User \ Desktop \ database \ Jaffe \ train \ 8 \ up \ '; featurepath = 'C: \ Users \ User \ Desktop \ mydoc \ expressiondata \ Jaffe \ DCT \ TXT \ DCT-train-Up.txt '; for I = for J = mydir = Invalid filename,num2str( I .pdf, '-', num2str(j),'.jpg']; I = imread (mydir); disp (mydir); % I = I '; A = dct2 (I); % imwrite (,''); B = zigzag2dto1d (a); if flag = 1 matrix = B (); flag = 2; else matrix = [matrix; B ()]; end end dlmwrite (featurepath, matrix, '',); % disp (matrix); End