gaborxiaobo.rar_小波图像检索

% GaborTextureSegment: Texture Segmentation using Gabor Filters % % seg = GaborTextureSegment(I,K,Sx,Sy,F,Theta,FUN,Wx,Wy,sigma) % % Input and output arguments ([]'s are optional): % I (matrix) of size NxM: Input Image of size NxM. % K (scalar): The # of expected segments. A param for k-means. % % gamma (scalar): The spatial aspect ratio, x to y. （x与y的空间比率） % Lambda (vector): The wavelengthes of the sinusoidal functions（正弦函数的波长）. % b (scalar): The spatial frequency band-width（空间频率带宽） (in octaves) % Theta (vector): The orientations of the gabor filters（高斯滤波的方位）. % phi (scalar): The phase offset（相位偏移量）. 0 is real part （实部）of Gabor filter or % even-symmetric（偶对称）, pi/2 is imaginary part （虚部）of Gabor filter or % odd-symmetric（奇对称）. % shape (string): shape option for conv2 to apply Gabor filter to image. % 'same' or 'valid' or 'full'. See help conv2. % (valid:Returns only those parts of the convolution that are computed without the zero-padded edges. % Using this option, C has size [ma-mb+1,na-nb+1] when all(size(A) >= size(B)). % Otherwise conv2 returns [].) % Note: sigma (scalar): The spread of Gabor filter or the standard % deviation of Gaussian is automatically computed from lambda and b. % （sigma是高斯标准差，由lambda和b计算而来） % seg (matrix) of size NxM: The segmented image. Each element has its % partition（分割） id. % % Author : Naotoshi Seo, sonots@umd.edu % Date : Oct, 2006 function seg = GaborTextureSegment(I, K, gamma, Lambda, b, Theta, phi, shape) [nRow, nCol, C] = size(I); % Step 1. Gabor Filter bank i = 0; for lambda = Lambda for theta = Theta i = i + 1; D = gabor2(I, gamma, lambda, b, theta, phi, shape); % Normalize into [0, 1] D = D - min(reshape(D, [], 1)); D = D / max(reshape(D, [], 1)); %figure; imshow(uint8(D * 255)); % Adjust image size to the smallest size if 'valid' (Cut off) if (isequal(shape, 'valid') && i >= 2) [nRow, nCol, C] = size(O(:, :, i-1)); [Nr, Nc, C] = size(D); DNr = (Nr - nRow)/2; DNc = (Nc - nCol)/2; D = D(1+floor(DNr):Nr-ceil(DNr), 1+floor(DNc):Nc-ceil(DNc)); end O(:, :, i) = D; end end [nRow, nCol, N] = size(O); % Step 2. Energy (Feature Extraction) % Step 2-1. Nonlinearity for i=1:N D = O(:, :, i); alpha = 1; D = tanh(double(O(:, :, i)) .* alpha); % Eq. (3). Input is [0, 1] % Normalize into [0, 1] although output of tanh is originally [0, 1] D = D - min(reshape(D, [], 1)); D = D / max(reshape(D, [], 1)); %figure; imshow(uint8(D * 255)); O(:, :, i) = D; end % Step 2-2. Smoothing for i=1:N D = O(:, :, i); lambda = Lambda(floor((i-1)/length(Theta))+1); % (1) constant % sigma = 5; % (2) Use lambda. 0.5 * lambda should be near equal to gabor's sigma % sigma = .5 * lambda; % (3). Use gabor's sigma sigma = (1 / pi) * sqrt(log(2)/2) * (2^b+1) / (2^b-1) * lambda; sigma = 3 * sigma; D = gauss2(D, sigma, shape); % Instead of Eq, (4) Avg filter % Normalize into [0, 1] D = D - min(reshape(D, [], 1)); D = D / max(reshape(D, [], 1)); %figure; imshow(uint8(D * 255)); % Adjust image size to the smallest size if 'valid' (Cut off) if (isequal(shape, 'valid') && i >= 2) [nRow, nCol, C] = size(P(:, :, i-1)); [Nr, Nc, C] = size(D); DNr = (Nr - nRow)/2; DNc = (Nc - nCol)/2; D = D(1+floor(DNr):Nr-ceil(DNr), 1+floor(DNc):Nc-ceil(DNc)); end P(:, :, i) = D; end O = P; clear P; [nRow, nCol, N] = size(O); % Step 3. Clustering % Step 3-1. Adding coordinates information to involves adjacency for i=1:nRow for j=1:nCol O(i, j, N+1) = i / nRow; % [0, 1] O(i, j, N+2) = j / nCol; end end % Step 3-2. Clustering data = reshape(O, [], size(O, 3)); % (# of data) x (# of dimensions) [cluster, codebook] = kmeans_light(data, K); %stopIter的默认值为0.05 seg = reshape(cluster, nRow, nCol, 1); % 2D end