如何在每个超像素上提取Gabor纹理特征？（含Matlab代码）

Hey there! Let's walk through how to extract Gabor features for each superpixel and create independent variables for your clustered categories—building right off the work you've already completed.

First, a quick recap of your existing work #

You've successfully run superpixel segmentation and clustered superpixels based on mean color. Here's your code and results organized clearly:

Superpixel Segmentation Code #

A = imread('kobi.png'); 
[L,N] = superpixels(A,5); 
figure 
BW = boundarymask(L); 
figure;imshow(imoverlay(A,BW,'cyan'),'InitialMagnification',67)

Superpixel Segmentation Result:

Superpixel Clustering Code #

idx=label2idx(L); 
meanColor = zeros(N,3); 
[m,n] = size(L); 
for i = 1:N 
    meanColor(i,1) = mean(A(idx{i})); 
    meanColor(i,2) = mean(A(idx{i}+m*n)); 
    meanColor(i,3) = mean(A(idx{i}+2*m*n)); 
end 
numColors = 6; 
[pidx,cmap] = kmeans(meanColor,numColors,'replicates',2); 
cmap = lab2rgb(cmap); 
Lout = zeros(size(A,1),size(A,2)); 
for i = 1:N 
    Lout(idx{i}) = pidx(i); 
end 
imshow(label2rgb(Lout))

Clustering Result:

Extracting Gabor Features for Each Superpixel #

Gabor filters are perfect for capturing texture across different scales and orientations. Here's how to compute them for each of your superpixels:

1. Generate a Gabor Filter Bank #

First, create a set of multi-scale, multi-directional Gabor filters using MATLAB's built-in gabor function:

% Define parameters: 4 scales, 6 orientations (adjust these as needed!)
num_scales = 4;
num_orientations = 6;
gabor_filters = gabor([0.5 1 2 4], 0:pi/num_orientations:pi-pi/num_orientations);

2. Compute Gabor Responses for the Image #

Convert your RGB image to grayscale (Gabor works best with single-channel data) and calculate responses for each filter:

gray_img = rgb2gray(A);
gabor_responses = imgaborfilt(gray_img, gabor_filters);

3. Extract Statistical Features per Superpixel #

Use the label2idx(L) you already used to get pixel indices for each superpixel, then pull out key stats (mean and variance are standard choices) from the Gabor responses:

superpixel_idx = label2idx(L);
total_filters = num_scales * num_orientations;
% Each superpixel gets 2 features per filter (mean + variance)
gabor_features = zeros(N, total_filters * 2);

for i = 1:N
    pix_indices = superpixel_idx{i}; % Pixels in current superpixel
    for j = 1:total_filters
        resp = gabor_responses(:,:,j);
        resp_pixels = resp(pix_indices);
        % Store mean and variance for this filter
        gabor_features(i, 2*j-1) = mean(resp_pixels);
        gabor_features(i, 2*j) = var(resp_pixels);
    end
end

Now gabor_features is an N-row matrix where each row holds the full texture feature set for one superpixel.

Creating Independent Variables for Each Clustered Category #

Based on your clustering results (pidx for superpixel labels, Lout for image-level labels), here are three practical ways to create independent variables:

Option 1: Category Mask Variables #

Generate a binary mask for each category—pixels in the category are 1, others are 0. Great for visualization or region-specific analysis:

num_classes = numColors; % You used 6 clusters
category_masks = cell(num_classes, 1);

for c = 1:num_classes
    category_masks{c} = (Lout == c); % Mask for category c
end

Call category_masks{3} to get the mask for your 3rd cluster anytime.

Option 2: One-Hot Encoded Variables #

Convert superpixel category labels into one-hot vectors—ideal for feeding into machine learning models:

one_hot_features = zeros(N, num_classes);
for i = 1:N
    one_hot_features(i, pidx(i)) = 1;
end

For example, if superpixel 5 belongs to cluster 2, one_hot_features(5,:) will be [0 1 0 0 0 0].

Option 3: Structured Category Info #

Store detailed info for each category (like which superpixels belong to it, average features) in a struct for easy access:

category_info = struct();
for c = 1:num_classes
    % Get all superpixels in this category
    class_superpixels = find(pidx == c);
    % Store metadata and aggregated features
    category_info.(['cluster', num2str(c)]) = struct(...
        'superpixel_indices', class_superpixels, ...
        'avg_gabor_features', mean(gabor_features(class_superpixels, :)) ...
    );
end

You can then pull up details like category_info.cluster1.avg_gabor_features to get the average Gabor features for all superpixels in cluster 1.

内容的提问来源于stack exchange，提问作者Aadnan Farooq A