在学习了斯坦福机器学习公开课中关于神经网络的章节之后，一直想自己实现一个神经网络，但是在没有对照的情况下，实在是无法验证自己写的程序对不对，因此，必须先找一个已经写好的神经网络来进行对比实验，刚好OpenCV中就有神经网络，于是，先使用OpenCV自带的神经网络对MNIST手写字体进行识别。
在OpenCV中，使用神经网络需要先设置一些关于神经网络的参数。
OpenCV中的神经网络是由类CvANN_MLP定义的，可以使用CvANN_MLP的构造函数或create函数产生一个神经网络，在产生神经网络的过程中，需要的参数如下：
因此，需要先设置神经网络每一层的神经元的个数，包括输入层和输出层，神经网络的激活函数，以及激活函数中对应的参数的值。
在这里，我主要选择sigmoid函数，sigmoid函数的参数都设置为1.
接下来，看一下神经网络的训练函数:
从函数的参数列表中可以看出，训练的过程需要一个CvANN_MLP_TrainParams类型的训练参数。
CvANN_MLP_TrainParams类型的说明如下：
我主要选择的是BP算法，因为另外一个算法没有研究过。
在了解了上面的需要设置的参数以后，就可以开始使用OpenCV的神经网络了。具体代码如下：
#include <opencv2/opencv.hpp>
#include <iostream>
#include "NeuralNetworksFunctions.h"
/**
* @brief NeuralNetworksTraing Training the neural networks
* @param NeuralNetworks The neural network
* @param InputMat Floating-point matrix of input vectors, one vector
* per row.
* @param OutputMat Floating-point matrix of the corresponding output
* vectors, one vector per row.
* @param MaxIte The number of the iteration
*
*
* @author sheng
* @version 1.0.0
* @date 2014-04-17
*
* @histroy <author> <date> <version> <description>
* sheng 2014-04-08 1.0.0 build the function
*
*/
void NeuralNetworksTraing(CvANN_MLP& NeuralNetworks, const cv::Mat& InputMat,
const cv::Mat& OutputMat, int MaxIte)
{
// Network architecture
std::vector<int> LayerSizes;
LayerSizes.push_back(InputMat.cols); // input layer
LayerSizes.push_back(1000); // hidden layer has 1000 neurons
LayerSizes.push_back(OutputMat.cols); // output layer
// Activate function
int ActivateFunc = CvANN_MLP::SIGMOID_SYM;
double Alpha = 1;
double Beta = 1;
// create the network
NeuralNetworks.create(cv::Mat(LayerSizes), ActivateFunc, Alpha, Beta);
// Training Params
CvANN_MLP_TrainParams TrainParams;
TrainParams.train_method = CvANN_MLP_TrainParams::BACKPROP;
TrainParams.bp_dw_scale = 0.0001;
TrainParams.bp_moment_scale = 0;
// iteration number
CvTermCriteria TermCrlt;
TermCrlt.type = CV_TERMCRIT_ITER | CV_TERMCRIT_EPS;
TermCrlt.epsilon = 0.0001f;
TermCrlt.max_iter = MaxIte;
TrainParams.term_crit = TermCrlt;
// Training the networks
NeuralNetworks.train(InputMat, OutputMat, cv::Mat(), cv::Mat(), TrainParams);
}
/**
* @brief NeuralNetworksPredict
* @param NeuralNetworks
* @param Input
* @param Output
*/
void NeuralNetworksPredict(const CvANN_MLP& NeuralNetworks, const cv::Mat& Input,
cv::Mat& Output)
{
// Neural network predict
cv::Mat OutputVector;
NeuralNetworks.predict(Input, OutputVector);
// change the output vector
DecodeOutputVector(OutputVector, Output, OutputVector.cols);
}
下面是测试代码：
#include "NeuralNetworksFunctions.h"
#include "MNIST.h"
#include "timer.h"
void Test_NeuralNetwork()
{
// prepare the training data
std::string TrainingImageFileName =
"train-images.idx3-ubyte";
cv::Mat TrainingImages = ReadImages(TrainingImageFileName);
cv::Mat FloatTrainingImages;
ConvertToFloatMat(TrainingImages, FloatTrainingImages);
// normalizing the training samples
FloatTrainingImages = (255 - FloatTrainingImages) / 255;
// prepare the training label
std::string TrainingLabelFileName =
"train-labels.idx1-ubyte";
cv::Mat TrainingLabels = ReadLabels(TrainingLabelFileName);
cv::Mat TrainingLabelVector;
EncodeOutputVector(TrainingLabels, TrainingLabelVector, 10);
// defining the network
CvANN_MLP Networks;
// The number of iteration
int MaxIte = 2;
// training
Utility::Timer NetworksTimer;
std::cout << "Training is started." << std::endl;
NetworksTimer.Start();
NeuralNetworksTraing(Networks, FloatTrainingImages, TrainingLabelVector,
MaxIte);
NetworksTimer.Finish();
std::cout << "Training is end." << std::endl;
std::cout << "The training time is " << NetworksTimer.GetDuration()
<< std::endl;
// save the networks
Networks.save("NerualNetworks-ite=2-1000hidden.xml");
// prapare the testing data
std::string TestingImageFileName =
"t10k-images.idx3-ubyte";
cv::Mat TestingImages = ReadImages(TestingImageFileName);
cv::Mat FloatTestingImages;
ConvertToFloatMat(TestingImages, FloatTestingImages);
// normalizing the testing samples
FloatTestingImages = (255 - FloatTestingImages) / 255;
// predicting
cv::Mat NetworkOutput;
NeuralNetworksPredict(Networks, FloatTestingImages, NetworkOutput);
// the actural output of the testing samples
std::string TestingLabelFileName =
"t10k-labels.idx1-ubyte";
cv::Mat TestingLabels = ReadLabels(TestingLabelFileName);
// calculating the number of the correct testing samples
int NumberOfCorrect = 0;
for (cv::MatIterator_<uchar> NetworkOutputIte = NetworkOutput.begin<uchar>(),
ActuralOutputIte = TestingLabels.begin<uchar>();
NetworkOutputIte != NetworkOutput.end<uchar>();
NetworkOutputIte++, ActuralOutputIte++)
{
if ((*NetworkOutputIte) == (*ActuralOutputIte))
{
NumberOfCorrect++;
}
}
float Rate = 0;
if (NetworkOutput.rows == 0)
{
std::cout << "The testing samples is zero." << std::endl;
}
else
{
Rate = (float)(NumberOfCorrect) / NetworkOutput.rows;
}
std::cout << "The number of correct is " << NumberOfCorrect << std::endl;
std::cout << "The number of the testing is " << NetworkOutput.rows << std::endl;
std::cout << "The correct of the network is " << Rate << std::endl;
}
在设置了使用BP算法，迭代次数为2，隐藏层神经元为1000的情况下，得到下面的结果
实验的效果还好，有90%的正确率。