本文利用560个(14个汉字,每字40)车牌汉字样本构成knn数据集,140个汉字样本作为测试集(每字10);
总结:0、读取样本时先转为灰度图,通道数为1
1、所用样本的宽高为28*28,将样本拉成28*28=784的向量,再塞进样本矩阵中(一个行数为样本总数600,列数为样本特征数784的矩阵),标签矩阵行数于样本矩阵行数必须相等,列数为1(600*1)
2、样本和标签矩阵数据类型必须为CV_32F,且通道数要为1,否则会报错
3、在构造标签矩阵时,尝试了下图中的方法,但标签矩阵内数据并不是目标数据(0或乱码),于是改用push_back的方法,输出矩阵正常
法一(有问题)
法二(没问题)
标签矩阵输出结果:
预测结果与对应标签:
最终预测准确率为87%,测试样本只用了140,其中粤和陕正确识别率较低。
void MainWindow::knnTrain(cv::Mat &src,bool ChiMoe)
{
using namespace std;
using namespace cv;
using namespace cv::ml;
string province[15] = {"云","皖","苏","辽","闽","黑","京","川","沪","浙","湘","粤","陕","鲁"};
Mat traindata,testData,trainlabel,testlabel;
GenerateDataSet(traindata,testData,trainlabel,testlabel,ChiMoe);//生成训练和测试的数据和标签,具体实现见下文
cout << format(testlabel,Formatter::FMT_NUMPY) << endl;//输出标签,开始时使用copyto函数构造标签矩阵,但是标签矩阵内数据都是0,改为push_back数据正常
trainlabel.convertTo(trainlabel,CV_32FC1);//样本和标签需为CV_32F
testlabel.convertTo(testlabel,CV_32FC1);
//knn模型初始化
cv::Ptr<cv::ml::KNearest> KNN = cv::ml::KNearest::create();
KNN->setDefaultK(6);
KNN->setIsClassifier(true);
KNN->setAlgorithmType(cv::ml::KNearest::BRUTE_FORCE);
KNN->train(traindata,cv::ml::ROW_SAMPLE,trainlabel);
Mat result;
KNN->findNearest(testData,6,result);
int t=0,f=0;//记录正确与错误数量
for(int i=0;i<140;++i){
int predict = int(result.at<float>(i));
int actual = int(testlabel.at<float>(i));
if(predict==actual){
qDebug() << "predict: " << predict << "actual: " << actual << "Y" << endl;
t++;//正确识别数
}
else{
qDebug() << "predict: " << predict << "actual: " << actual << "X" << endl;
f++;//错误识别数
}
}
qDebug() << "The Correct Rate:" << float(t)/float(140) << endl;//输出识别正确率
}
数据集生成函数:
void MainWindow::GenerateDataSet(cv::Mat &trainData, cv::Mat &testData, cv::Mat &trainLabel, cv::Mat &testLabel,bool ChiMode)
{
using namespace std;
using namespace cv;
string province[14] = {"云","皖","苏","辽","闽","黑","京","川","沪","浙","湘","粤","陕","鲁"};
Mat data,label;
for(size_t i=0;i<14;++i){
for(int j=1;j<=40;++j){
string path = "/home/ghoson-x/Desktop/qt/CarImage/data2/train_28_28/";
path.append(province[i]).append("/").append(province[i]).append("_").append((QString::number(j)).toStdString()).append(".jpg");
Mat tmp = imread(path,IMREAD_GRAYSCALE);
if(tmp.empty()){
break;
}
tmp.convertTo(tmp,CV_32FC1);
trainData.push_back(tmp.reshape(0,1));//图片拉成一条784的向量
}
Mat tmp_label = Mat::ones(40,1,CV_32FC1);//各样本数为40,总计600
tmp_label = tmp_label * i;
trainLabel.push_back(tmp_label);
//生成测试标签
for(int k=1;k<=10;++k){
string path = "/home/ghoson-x/Desktop/qt/CarImage/data2/test_28_28/";
path.append(province[i]).append("/").append(province[i]).append("_").append((QString::number(k)).toStdString()).append(".jpg");
Mat tmp = imread(path,IMREAD_GRAYSCALE);
if(tmp.empty()){
break;
}
tmp.convertTo(tmp,CV_32FC1);
testData.push_back(tmp.reshape(0,1));
}
Mat test_label = Mat::ones(10,1,CV_32FC1);
test_label = test_label * i;
cout << format(test_label,Formatter::FMT_NUMPY) << endl;
testLabel.push_back(test_label);
}
}