颜色特征提取（三）------颜色聚合向量

Pass[9]提出了图像的颜色聚合向量（color coherence vector）。它是颜色直方图的一种演变，其核心思想是将属于直方图每一个bin的像素进行分为两部分：如果该bin内的某些像素所占据的连续区域的面积大于给定的阈值，则该区域内的像素作为聚合像素，否则作为非聚合像素。假设αi与βi分别代表直方图的第i个bin中聚合像素和非聚合像素的数量，图像的颜色聚合向量可以表达为<(α1, β1), (α2, β2), …, (αN, βN)>。而<α1+ β1, α2 + β2, …,
αN +βN > 就是该图像的颜色直方图。由于包含了颜色分布的空间信息，颜色聚合向量相比颜色直方图可以达到更好的检索效果。

MATLAB实现：

% CCV : This is the normal Color Coherence Vector
% 
% ICCV : This is an Improved Color Coherence Vector by adding max coherent pixels’ spatial information without affecting the performance
% 
% Parallel implementation based on these papers :
% 
% 1) Comparing Images Using Color Coherence Vectors (1996) - http://goo.gl/LkWkbi -
% 
% 2) An Improved Color Coherence Vector - http://goo.gl/FjXHje -
% getCCV and getICCV function take an image and return the Color Coherence Vector that describe this Image
function CCV = getCCV(img,coherentPrec, numberOfColors)
    if ~exist('coherentPrec','var')
        coherentPrec = 1;
    end
    if ~exist('numberOfColors','var')
        numberOfColors = 27;
    end
    CCV = zeros(2,numberOfColors);
    
    Gaus = fspecial('gaussian',[5 5],2);
    img = imfilter(img,Gaus,'same');
    
    [img, updNumOfPix]= discretizeColors(img,numberOfColors);
    
    imgSize = (size(img,1)*size(img,2));
    thresh = int32((coherentPrec/100) *imgSize);
    
    parfor i=0:updNumOfPix-1
        BW = img==i;
        CC = bwconncomp(BW);
        compsSize = cellfun(@numel,CC.PixelIdxList);
        incoherent = sum(compsSize(compsSize>=thresh));
        CCV(:,i+1) = [incoherent; ...
            sum(compsSize) - incoherent];
    end
end

function ICCV = getICCV(img,coherentPrec, numberOfColors)
    if ~exist('coherentPrec','var')
        coherentPrec = 1;
    end
    if ~exist('numberOfColors','var')
        numberOfColors = 27;
    end
    ICCV = zeros(4,numberOfColors);
    
    Gaus = fspecial('gaussian',[5 5],2);
    img = imfilter(img,Gaus,'same');
    
    [img, updNumOfPix]= discretizeColors(img,numberOfColors);
    
    imgSize = (size(img,1)*size(img,2));
    thresh = int32((coherentPrec/100) *imgSize);
    
    for i=0:updNumOfPix-1
        BW = img==i;
        CC = bwconncomp(BW);
        compsSize = cellfun(@numel,CC.PixelIdxList);
        [~,idx] = max(compsSize);
        if isempty(idx)==0
            [subI, subJ] = ind2sub(size(img),CC.PixelIdxList{idx});
            meanPos = uint32(mean([subI subJ],1));
        else
            meanPos = [0 0];
        end
        incoherent = sum(compsSize(compsSize>=thresh));
        ICCV(:,i+1) = [incoherent; ...
            sum(compsSize) - incoherent;meanPos'];
    end
end
function [oneChannel, updatedNumC] = discretizeColors(img,numColors)

width = size(img,2);
height = size(img,1);
oneChannel = zeros(height,width);

% We have 3 channels. For each channel we have V unique values. 
% So we want to find the value of V given that V x V x V ~= numColors
numOfBins = floor(pow2(log2(numColors)/3));
numOfBinsSQ = numOfBins*numOfBins;
img = floor((img/(256/numOfBins)));
for i=1:height
    for j=1:width
        oneChannel(i,j) = img(i,j,1)*numOfBinsSQ ...
            + img(i,j,2)*numOfBins + img(i,j,3);
    end
end
updatedNumC = power(numOfBins,3);

end

分别对一张图片得到CCV和ICCV特征，然后进行后续的处理