基于K近邻（KNN）算法在MATLAB中实现人脸识别

一、代码实现

1. 数据预处理

function processedImg = preprocess(img)
    % 灰度化
    grayImg = rgb2gray(img);
    % 直方图均衡化
    equalizedImg = histeq(grayImg);
    % 归一化尺寸
    processedImg = imresize(equalizedImg, [100, 100]);
end

2. 特征提取（PCA降维）

function features = extractFeatures(images)
    % 将图像矩阵转换为特征向量
    [numSamples, height, width] = size(images);
    imgVectors = double(reshape(images, numSamples, height*width));
    
    % 计算均值脸
    meanFace = mean(imgVectors, 1);
    centered = imgVectors - meanFace;
    
    % 使用SVD计算特征向量
    [U, ~, ~] = svd(centered' * centered);
    eigFaces = U(:, 1:50); % 取前50个主成分
    
    % 投影到特征空间
    features = centered * eigFaces;
end

3. KNN分类器训练

% 加载ORL数据集（示例）
load('orl_faces.mat'); % 包含trainData/trainLabels和testData/testLabels

% 预处理数据
trainDataProc = arrayfun(@(i) preprocess(trainData(:,:,i)), 1:size(trainData,3), 'UniformOutput', false);
testDataProc = arrayfun(@(i) preprocess(testData(:,:,i)), 1:size(testData,3), 'UniformOutput', false);

% 提取特征
trainFeatures = extractFeatures(trainDataProc);
testFeatures = extractFeatures(testDataProc);

% 构建KNN模型
k = 5; % 近邻数
mdl = fitcknn(trainFeatures, trainLabels, 'NumNeighbors', k, 'Distance', 'euclidean');

4. 人脸识别测试

% 预测测试集
predictedLabels = predict(mdl, testFeatures);

% 计算识别率
accuracy = sum(predictedLabels == testLabels) / numel(testLabels);
disp(['识别准确率: ', num2str(accuracy*100), '%']);

---

二、优化

1. 多尺度特征提取

% 使用图像金字塔增强特征不变性
pyramidLevels = 3;
for level = 1:pyramidLevels
    scale = 2^(-level);
    scaledImg = imresize(img, scale);
    % 提取多尺度特征并融合
end

2. 动态K值选择

% 交叉验证选择最佳K值
cv = cvpartition(size(trainFeatures,1),'KFold',5);
kValues = 1:5;
bestAcc = 0;

for k = kValues
    cvAcc = crossval(@(X,Y) sum(predict(fitcknn(X,Y,k),Y) == Y)/numel(Y), ...
                     trainFeatures, trainLabels, 'partition', cv);
    if cvAcc > bestAcc
        bestK = k;
        bestAcc = cvAcc;
    end
end
disp(['最佳K值: ', num2str(bestK)]);

3. 特征空间优化

% 使用LDA替代PCA进行特征降维
[coeff, score, ~] = lda(trainFeatures, trainLabels);
projectedFeatures = score(:,1:10); % 降维到10维

---

三、工程实现建议

1. GPU加速

% 启用并行计算
parpool('local', 4);
gpuFeatures = gpuArray(trainFeatures);

2. 实时处理优化

% 视频流处理
vid = videoinput('winvideo', 1, 'YUY2_640x480');
vid.FramesPerTrigger = Inf;

while isvalid(vid)
    frame = getsnapshot(vid);
    processed = preprocess(frame);
    feature = extractFeatures(processed);
    label = predict(mdl, feature);
    imshow(frame);
    title(['识别结果: ', num2str(label)]);
    drawnow;
end

---

四、典型应用场景

1. 门禁系统

% 实时人脸匹配
databaseFeatures = load('face_database.mat'); % 预存特征库
[minDist, idx] = min(pdist2(testFeature, databaseFeatures));
if minDist < 0.5
    disp(['欢迎 ', databaseLabels(idx)]);
else
    disp('访问拒绝');
end

2. 移动端部署

% 模型轻量化
compressedModel = compact(mdl);
save('knn_model.mat', 'compressedModel');

参考代码 采用模式识别算法在matlab中利用近邻法进行人脸识别 www.youwenfan.com/contentcni/65217.html

五、常见问题解决方案

1. 光照变化

% 自适应直方图均衡化
img = adapthisteq(grayImg, 'ClipLimit', 0.02);

2. 姿态变化

% 使用仿射变换对齐人脸
tform = estimateGeometricTransform(eyeCorners, faceCorners, 'affine');
alignedImg = imwarp(faceImg, tform);

3. 误识别问题

% 引入置信度阈值
[~, scores] = predict(mdl, testFeature);
if max(scores) < 0.7
    label = '未知';
end