MATLAB实现手写体0-9数字识别

一、系统架构设计

graph TD
    A[数据采集] --> B(预处理)
    B --> C{特征提取}
    C --> D[传统方法]
    C --> E[深度学习]
    D --> F[KNN/SVM分类]
    E --> G[CNN分类]
    F --> H[识别结果]
    G --> H

二、核心实现步骤

1. 数据准备与预处理

% 加载MNIST数据集（MATLAB内置）
[XTrain, YTrain] = digitTrain4DArrayData;  % 60,000训练样本
[XTest, YTest] = digitTest4DArrayData;     % 10,000测试样本

% 数据归一化
XTrain = double(XTrain)/255;
XTest = double(XTest)/255;

% 转换为图像矩阵
trainImages = reshape(XTrain, 28, 28, 1, []);
testImages = reshape(XTest, 28, 28, 1, []);

2. 传统机器学习方法

2.1 特征提取（HOG特征）

% 提取HOG特征
hogFeatureSize = 36;  % 9 bins × 4 cells
trainFeatures = extractHOGFeatures(trainImages, 'CellSize', [7 7]);
testFeatures = extractHOGFeatures(testImages, 'CellSize', [7 7]);

2.2 分类器训练

% KNN分类器
mdlKNN = fitcknn(trainFeatures, YTrain, 'NumNeighbors', 5, 'Standardize', true);

% SVM分类器
mdlSVM = fitcecoc(trainFeatures, YTrain, 'Learners', 'linear', 'Coding', 'onevsall');

三、深度学习方法（CNN）

1. 网络结构设计

layers = [
    imageInputLayer([28 28 1])  % 输入层
    
    convolution2dLayer(3, 8, 'Padding', 'same')  % 卷积层1
    batchNormalizationLayer()
    reluLayer()
    maxPooling2dLayer(2, 'Stride', 2)
    
    convolution2dLayer(3, 16, 'Padding', 'same') % 卷积层2
    batchNormalizationLayer()
    reluLayer()
    maxPooling2dLayer(2, 'Stride', 2)
    
    fullyConnectedLayer(64)        % 全连接层
    reluLayer()
    
    fullyConnectedLayer(10)        % 输出层
    softmaxLayer()
    classificationLayer()];        % 分类层

2. 模型训练

options = trainingOptions('adam',...
    'MaxEpochs', 15,...
    'MiniBatchSize', 64,...
    'InitialLearnRate', 0.001,...
    'Shuffle', 'every-epoch',...
    'Verbose', false);

net = trainNetwork(trainImages, YTrain, layers, options);

四、性能评估

1. 测试集验证

predictedLabels = classify(net, testImages);
accuracy = sum(predictedLabels == YTest)/numel(YTest);
fprintf('测试集准确率: %.2f%%
', accuracy*100);

2. 混淆矩阵分析

plotconfusion(YTest, predictedLabels);
title('混淆矩阵分析');

五、优化

1. 数据增强

augmentedTrainingSet = imageDataAugmenter(...
    'RandRotation', [-10 10],...
    'RandXReflection', true,...
    'RandYReflection', true);

augmentedData = augmentedTrainingSet(XTrain);

2. 迁移学习

net = alexnet;
lgraph = layerGraph(net);
newFCLayer = fullyConnectedLayer(10, 'Name', 'new_fc');
newClassLayer = classificationLayer('Name', 'new_classoutput');
lgraph = replaceLayer(lgraph, 'fc1000', newFCLayer);
lgraph = replaceLayer(lgraph, 'ClassificationLayer_fc1000', newClassLayer);

net = trainNetwork(XTrain, YTrain, lgraph, options);

六、工程实现方案

1. 交互式GUI设计（App Designer）

% 创建手写输入界面
fig = uifigure('Name', '手写数字识别');
ax = uiaxes(fig, 'Position', [0.1 0.3 0.8 0.6]);
btn = uibutton(fig, 'Text', '识别', 'Position', [0.4 0.1 0.2 0.1],...
    'ButtonPushedFcn', @(btn,event) recognizeDigit());

% 实时绘制手写轨迹
ax.XTick = [];
ax.YTick = [];
ax.XColor = 'w';
ax.YColor = 'w';
hold(ax, 'on');
currentPoint = [0,0];

% 鼠标事件回调
function mouseMoved(~, event)
    currentPoint = [event.Position(1), event.Position(2)];
    plot(ax, currentPoint(1), currentPoint(2), 'r.', 'MarkerSize', 20);
end

2. 移动端部署

% 生成Android应用
mcc -m HandwritingRecognition -a assets -d compiled;

七、完整代码仓库

% 主程序文件: digitRecognition.m
% 数据加载: loadMNISTData.m
% 特征提取: extractFeatures.m
% 模型训练: trainModels.m
% 测试评估: evaluatePerformance.m

八、参考

MathWorks官方示例: Digit Classification Using HOG Features https://ww2.mathworks.cn/help/deeplearning/examples/digit-classification-using-hog-features.html
参考代码识别手写体0-9的数字 www.youwenfan.com/contentcso/78288.html
周志华《机器学习》第5章支持向量机