在opencv3中的机器学习算法练习：对OCR进行分类

OCR （Optical Character Recognition，光学字符识别），我们这个练习就是对OCR英文字母进行识别。得到一张OCR图片后，提取出字符相关的ROI图像，并且大小归一化，整个图像的像素值序列可以直接作为特征。但直接将整个图像作为特征数据维度太高，计算量太大，所以也可以进行一些降维处理，减少输入的数据量。

处理过程一般这样：先对原图像进行裁剪，得到字符的ROI图像，二值化。然后将图像分块，统计每个小块中非0像素的个数，这样就形成了一个较小的矩阵，这矩阵就是新的特征了。opencv为我们提供了一些这样的数据，放在

opencvsourcessamplesdataletter-recognition.data

这个文件里，打开看看：

每一行代表一个样本。第一列大写的字母，就是标注，随后的16列就是该字母的特征向量。这个文件中总共有20000行样本，共分类26类（26个字母）。

我们将这些数据读取出来后，分成两部分，第一部分16000个样本作为训练样本，训练出分类器后，对这16000个训练数据和余下的4000个数据分别进行测试，得到训练精度和测试精度。其中adaboost比较特殊一点，训练和测试样本各为10000.

完整代码为：

#include "stdafx.h"
#include "opencv2opencv.hpp"
#include <iostream>
using namespace std;
using namespace cv;
using namespace cv::ml;

// 读取文件数据
bool read_num_class_data(const string& filename, int var_count,Mat* _data, Mat* _responses)
{
    const int M = 1024;
    char buf[M + 2];

    Mat el_ptr(1, var_count, CV_32F);
    int i;
    vector<int> responses;

    _data->release();
    _responses->release();
    FILE *f;
    fopen_s(&f, filename.c_str(), "rt");
    if (!f)
    {
        cout << "Could not read the database " << filename << endl;
        return false;
    }

    for (;;)
    {
        char* ptr;
        if (!fgets(buf, M, f) || !strchr(buf, ','))
            break;
        responses.push_back((int)buf[0]);
        ptr = buf + 2;
        for (i = 0; i < var_count; i++)
        {
            int n = 0;
            sscanf_s(ptr, "%f%n", &el_ptr.at<float>(i), &n);
            ptr += n + 1;
        }
        if (i < var_count)
            break;
        _data->push_back(el_ptr);
    }
    fclose(f);
    Mat(responses).copyTo(*_responses);
    return true;
}


//准备训练数据
Ptr<TrainData> prepare_train_data(const Mat& data, const Mat& responses, int ntrain_samples)
{
    Mat sample_idx = Mat::zeros(1, data.rows, CV_8U);
    Mat train_samples = sample_idx.colRange(0, ntrain_samples);
    train_samples.setTo(Scalar::all(1));

    int nvars = data.cols;
    Mat var_type(nvars + 1, 1, CV_8U);
    var_type.setTo(Scalar::all(VAR_ORDERED));
    var_type.at<uchar>(nvars) = VAR_CATEGORICAL;

    return TrainData::create(data, ROW_SAMPLE, responses,
        noArray(), sample_idx, noArray(), var_type);
}

//设置迭代条件
inline TermCriteria TC(int iters, double eps)
{
    return TermCriteria(TermCriteria::MAX_ITER + (eps > 0 ? TermCriteria::EPS : 0), iters, eps);
}

//分类预测
void test_and_save_classifier(const Ptr<StatModel>& model,    const Mat& data, const Mat& responses,
    int ntrain_samples, int rdelta)
{
    int i, nsamples_all = data.rows;
    double train_hr = 0, test_hr = 0;

    // compute prediction error on train and test data
    for (i = 0; i < nsamples_all; i++)
    {
        Mat sample = data.row(i);

        float r = model->predict(sample);
        r = std::abs(r + rdelta - responses.at<int>(i)) <= FLT_EPSILON ? 1.f : 0.f;

        if (i < ntrain_samples)
            train_hr += r;
        else
            test_hr += r;
    }

    test_hr /= nsamples_all - ntrain_samples;
    train_hr = ntrain_samples > 0 ? train_hr / ntrain_samples : 1.;

    printf("Recognition rate: train = %.1f%%, test = %.1f%%
",
        train_hr*100., test_hr*100.);
}

//随机树分类
bool build_rtrees_classifier(const string& data_filename)
{
    Mat data;
    Mat responses;
    read_num_class_data(data_filename, 16, &data, &responses);

    int nsamples_all = data.rows;
    int ntrain_samples = (int)(nsamples_all*0.8);

    Ptr<RTrees> model;
    Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
    model = RTrees::create();
    model->setMaxDepth(10);
    model->setMinSampleCount(10);
    model->setRegressionAccuracy(0);
    model->setUseSurrogates(false);
    model->setMaxCategories(15);
    model->setPriors(Mat());
    model->setCalculateVarImportance(true);
    model->setActiveVarCount(4);
    model->setTermCriteria(TC(100, 0.01f));
    model->train(tdata);
    test_and_save_classifier(model, data, responses, ntrain_samples, 0);
    cout << "Number of trees: " << model->getRoots().size() << endl;

    // Print variable importance
    Mat var_importance = model->getVarImportance();
    if (!var_importance.empty())
    {
        double rt_imp_sum = sum(var_importance)[0];
        printf("var#	importance (in %%):
");
        int i, n = (int)var_importance.total();
        for (i = 0; i < n; i++)
            printf("%-2d	%-4.1f
", i, 100.f*var_importance.at<float>(i) / rt_imp_sum);
    }

    return true;
}

//adaboost分类
bool build_boost_classifier(const string& data_filename)
{
    const int class_count = 26;
    Mat data;
    Mat responses;
    Mat weak_responses;

    read_num_class_data(data_filename, 16, &data, &responses);
    int i, j, k;
    Ptr<Boost> model;

    int nsamples_all = data.rows;
    int ntrain_samples = (int)(nsamples_all*0.5);
    int var_count = data.cols;

    Mat new_data(ntrain_samples*class_count, var_count + 1, CV_32F);
    Mat new_responses(ntrain_samples*class_count, 1, CV_32S);

    for (i = 0; i < ntrain_samples; i++)
    {
        const float* data_row = data.ptr<float>(i);
        for (j = 0; j < class_count; j++)
        {
            float* new_data_row = (float*)new_data.ptr<float>(i*class_count + j);
            memcpy(new_data_row, data_row, var_count*sizeof(data_row[0]));
            new_data_row[var_count] = (float)j;
            new_responses.at<int>(i*class_count + j) = responses.at<int>(i) == j + 'A';
        }
    }

    Mat var_type(1, var_count + 2, CV_8U);
    var_type.setTo(Scalar::all(VAR_ORDERED));
    var_type.at<uchar>(var_count) = var_type.at<uchar>(var_count + 1) = VAR_CATEGORICAL;

    Ptr<TrainData> tdata = TrainData::create(new_data, ROW_SAMPLE, new_responses,
        noArray(), noArray(), noArray(), var_type);
    vector<double> priors(2);
    priors[0] = 1;
    priors[1] = 26;

    model = Boost::create();
    model->setBoostType(Boost::GENTLE);
    model->setWeakCount(100);
    model->setWeightTrimRate(0.95);
    model->setMaxDepth(5);
    model->setUseSurrogates(false);
    model->setPriors(Mat(priors));
    model->train(tdata);
    Mat temp_sample(1, var_count + 1, CV_32F);
    float* tptr = temp_sample.ptr<float>();

    // compute prediction error on train and test data
    double train_hr = 0, test_hr = 0;
    for (i = 0; i < nsamples_all; i++)
    {
        int best_class = 0;
        double max_sum = -DBL_MAX;
        const float* ptr = data.ptr<float>(i);
        for (k = 0; k < var_count; k++)
            tptr[k] = ptr[k];

        for (j = 0; j < class_count; j++)
        {
            tptr[var_count] = (float)j;
            float s = model->predict(temp_sample, noArray(), StatModel::RAW_OUTPUT);
            if (max_sum < s)
            {
                max_sum = s;
                best_class = j + 'A';
            }
        }

        double r = std::abs(best_class - responses.at<int>(i)) < FLT_EPSILON ? 1 : 0;
        if (i < ntrain_samples)
            train_hr += r;
        else
            test_hr += r;
    }

    test_hr /= nsamples_all - ntrain_samples;
    train_hr = ntrain_samples > 0 ? train_hr / ntrain_samples : 1.;
    printf("Recognition rate: train = %.1f%%, test = %.1f%%
",
        train_hr*100., test_hr*100.);

    cout << "Number of trees: " << model->getRoots().size() << endl;
    return true;
}

//多层感知机分类（ANN）
bool build_mlp_classifier(const string& data_filename)
{
    const int class_count = 26;
    Mat data;
    Mat responses;

    read_num_class_data(data_filename, 16, &data, &responses);
    Ptr<ANN_MLP> model;

    int nsamples_all = data.rows;
    int ntrain_samples = (int)(nsamples_all*0.8);
    Mat train_data = data.rowRange(0, ntrain_samples);
    Mat train_responses = Mat::zeros(ntrain_samples, class_count, CV_32F);

    // 1. unroll the responses
    cout << "Unrolling the responses...
";
    for (int i = 0; i < ntrain_samples; i++)
    {
        int cls_label = responses.at<int>(i) -'A';
        train_responses.at<float>(i, cls_label) = 1.f;
    }

    // 2. train classifier
    int layer_sz[] = { data.cols, 100, 100, class_count };
    int nlayers = (int)(sizeof(layer_sz) / sizeof(layer_sz[0]));
    Mat layer_sizes(1, nlayers, CV_32S, layer_sz);

#if 1
    int method = ANN_MLP::BACKPROP;
    double method_param = 0.001;
    int max_iter = 300;
#else
    int method = ANN_MLP::RPROP;
    double method_param = 0.1;
    int max_iter = 1000;
#endif

    Ptr<TrainData> tdata = TrainData::create(train_data, ROW_SAMPLE, train_responses);
    model = ANN_MLP::create();
    model->setLayerSizes(layer_sizes);
    model->setActivationFunction(ANN_MLP::SIGMOID_SYM, 0, 0);
    model->setTermCriteria(TC(max_iter, 0));
    model->setTrainMethod(method, method_param);
    model->train(tdata);
    return true;
}

//K最近邻分类
bool build_knearest_classifier(const string& data_filename, int K)
{
    Mat data;
    Mat responses;
    read_num_class_data(data_filename, 16, &data, &responses);
    int nsamples_all = data.rows;
    int ntrain_samples = (int)(nsamples_all*0.8);

    Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
    Ptr<KNearest> model = KNearest::create();
    model->setDefaultK(K);
    model->setIsClassifier(true);
    model->train(tdata);

    test_and_save_classifier(model, data, responses, ntrain_samples, 0);
    return true;
}

//贝叶斯分类
bool build_nbayes_classifier(const string& data_filename)
{
    Mat data;
    Mat responses;
    read_num_class_data(data_filename, 16, &data, &responses);

    int nsamples_all = data.rows;
    int ntrain_samples = (int)(nsamples_all*0.8);

    Ptr<NormalBayesClassifier> model;
    Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
    model = NormalBayesClassifier::create();
    model->train(tdata);

    test_and_save_classifier(model, data, responses, ntrain_samples, 0);
    return true;
}


//svm分类
bool build_svm_classifier(const string& data_filename)
{
    Mat data;
    Mat responses;
    read_num_class_data(data_filename, 16, &data, &responses);

    int nsamples_all = data.rows;
    int ntrain_samples = (int)(nsamples_all*0.8);

    Ptr<SVM> model;
    Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
    model = SVM::create();
    model->setType(SVM::C_SVC);
    model->setKernel(SVM::LINEAR);
    model->setC(1);
    model->train(tdata);

    test_and_save_classifier(model, data, responses, ntrain_samples, 0);
    return true;
}

int main()
{
    string data_filename = "E:/opencv/opencv/sources/samples/data/letter-recognition.data";  //字母数据

    cout << "svm分类：" << endl;
    build_svm_classifier(data_filename);

    cout << "贝叶斯分类：" << endl; 
    build_nbayes_classifier(data_filename);

    cout << "K最近邻分类：" << endl;
    build_knearest_classifier(data_filename,10);    

    cout << "随机树分类：" << endl;
    build_rtrees_classifier(data_filename);
    
    //cout << "adaboost分类：" << endl;
    //build_boost_classifier(data_filename);

    //cout << "ANN（多层感知机)分类：" << endl;
    //build_mlp_classifier(data_filename);
}

由于adaboost分类和 ann分类速度非常慢，因此我在main函数里把这两个分类注释掉了，大家有兴趣和时间可以测试一下。

结果：

从结果显示来看，测试的四种分类算法中，KNN（最近邻）分类精度是最高的。所以说，对ocr进行识别，还是用knn最好。