Got it, I’ve put together a clear, step-by-step MATLAB SVM classification example tailored for your 8-class object recognition task. This code uses synthetic data (super easy to swap out with your real target recognition datasets) and covers all the key parts you’ll need to learn: data setup, model training, prediction, performance checks, and even visualization. Let’s jump in:

Full Runable Code #

%% 8-Class SVM Classification for Object Recognition
clear; clc; close all;

% --------------------------
% 1. Generate Synthetic 8-Class Data (Replace with your real features)
% --------------------------
numClasses = 8;
samplesPerClass = 100;
featuresPerSample = 2; % Swap with your actual feature dimension (e.g., 128 for HOG)

% Create distinct clusters for each class
data = [];
labels = [];
for i = 1:numClasses
    % Shift clusters to make classes distinguishable
    classData = randn(samplesPerClass, featuresPerSample) + (i-1)*2;
    data = [data; classData];
    labels = [labels; repmat(categorical(['Class ', num2str(i)]), samplesPerClass, 1)];
end

% Split into 70% training / 30% testing sets (per class to avoid bias)
rng(42); % Set seed for reproducible results
trainIdx = [];
testIdx = [];
for i = 1:numClasses
    classIdx = find(labels == categorical(['Class ', num2str(i)]));
    splitIdx = splitIndices(length(classIdx), 0.7);
    trainIdx = [trainIdx; classIdx(splitIdx(1))];
    testIdx = [testIdx; classIdx(splitIdx(2))];
end
trainData = data(trainIdx, :);
trainLabels = labels(trainIdx);
testData = data(testIdx, :);
testLabels = labels(testIdx);

% --------------------------
% 2. Train Multi-Class SVM Model
% --------------------------
% Use fitcecoc (Error-Correcting Output Codes) for multi-class SVM
% RBF kernel works well for non-linear data; use 'Linear' if your data is separable
svmModel = fitcecoc(trainData, trainLabels, ...
    'Learners', 'SVM', ...
    'KernelFunction', 'RBF', ...
    'KernelScale', 'auto', ...
    'BoxConstraint', 1);

% --------------------------
% 3. Predict on Test Set & Calculate Accuracy
% --------------------------
predictedLabels = predict(svmModel, testData);
accuracy = mean(predictedLabels == testLabels);
fprintf('Test Set Accuracy: %.2f%%\n', accuracy*100);

% --------------------------
% 4. Visualize Confusion Matrix
% --------------------------
figure;
confusionchart(testLabels, predictedLabels);
title('Confusion Matrix for 8-Class Object Recognition');
xlabel('Predicted Class');
ylabel('True Class');

% --------------------------
% 5. Plot Decision Boundaries (Only for 2D features)
% --------------------------
if featuresPerSample == 2
    figure;
    gscatter(trainData(:,1), trainData(:,2), trainLabels);
    hold on;
    % Create grid to map decision regions
    [x1, x2] = meshgrid(min(data(:,1)):0.1:max(data(:,1)), min(data(:,2)):0.1:max(data(:,2)));
    gridData = [x1(:), x2(:)];
    gridPred = predict(svmModel, gridData);
    % Overlay decision regions
    gscatter(gridData(:,1), gridData(:,2), gridPred, [], [], 0.1);
    hold off;
    title('SVM Decision Boundaries for 8 Classes');
    xlabel('Feature 1');
    ylabel('Feature 2');
    legend('Location', 'bestoutside');
end

Code Breakdown #

Let’s walk through what each part does so you can adapt it to your own data:

• Data Generation: I made 8 distinct clusters to simulate object features. You’ll want to replace this with your actual data—like HOG features extracted from your target images, or precomputed CNN embeddings.
• Train-Test Split: We split each class evenly to avoid bias, so the model learns from most samples and is tested on unseen ones. The rng(42) line ensures your results are reproducible.
• SVM Training: fitcecoc is MATLAB’s standard tool for multi-class SVM (since basic SVM is binary). We use an RBF kernel here, which works well for most real-world non-linear data. If your features are linearly separable, swap 'RBF' with 'Linear'.
• Prediction & Accuracy: We test the model on held-out data and calculate how often it gets the class right. The confusion matrix shows exactly where the model might be mixing up classes.
• Visualization: Only works for 2D features (as in this example). For high-dimensional features (common in object recognition), skip this part or use PCA to reduce dimensions for plotting.

Pro Tips for Your Object Recognition Task #

• Feature Extraction: Use MATLAB’s extractHOGFeatures to get robust image features, or import features you’ve generated with other tools (like OpenCV).
• Hyperparameter Tuning: To boost performance, use bayesopt to optimize parameters like KernelScale (for RBF) and BoxConstraint. This helps the model fit your data better.
• Handle Class Imbalance: If some object classes have way fewer samples than others, add 'ClassWeights', 'balanced' to the fitcecoc call to make the model prioritize underrepresented classes.

内容的提问来源于stack exchange，提问作者Hilal Güven