class_ann.h
2.68 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
//OpenCV
#include <opencv2/opencv.hpp>
#include "progress_thread.h"
extern size_t X, Y, B;
//function for training a random forest classifier
void train_ann(CvANN_MLP* ann, cv::Mat &trainF, cv::Mat &trainR, std::vector<int> hidden, int iters, double epsilon, bool VERBOSE = false){
//ANN criteria for termination
CvTermCriteria criter;
criter.max_iter = iters;
criter.epsilon = epsilon;
criter.type = CV_TERMCRIT_ITER | CV_TERMCRIT_EPS;
//ANN parameters
CvANN_MLP_TrainParams params;
params.train_method = CvANN_MLP_TrainParams::BACKPROP;
params.bp_dw_scale = 0.05f;
params.bp_moment_scale = 0.05f;
params.term_crit = criter; //termination criteria
cv::Mat layers = cv::Mat((int)hidden.size() + 2, 1, CV_32SC1); //create one input layer, one output layer, and n (specified) hidden layers
layers.row(0) = cv::Scalar(trainF.cols); //the first layer is based on the number of inputs
for(size_t l = 0; l < hidden.size(); l++){ //for each hidden layer
layers.row((int)l + 1) = cv::Scalar(hidden[l]); //specify the number of hidden nodes
}
layers.row((int)hidden.size() + 1) = cv::Scalar(trainR.cols); //the output layer is equal to the number of responses per pixel
std::cout<<"Creating an MLP with the structure: "<<std::endl<<"\t";
std::cout<<"(in) "<<layers.at<int>(0, 0);
for(size_t l = 1; l < layers.rows - 1; l++)
std::cout<<" "<<layers.at<int>((int)l, 0);
std::cout<<" "<<layers.at<int>(layers.rows-1, 0)<<" (out)"<<std::endl;
ann->create(layers); //create the neural network given the architecture
std::cout<<"Training..."<<std::endl;
ann->train(trainF, trainR, cv::Mat(), cv::Mat(), params); //call the ANN training algorithm
std::cout<<"done..."<<std::endl;
///// TEMPORARY - WILL BE DELETED IN THE ACTUAL IMPLEMENTATION ///////////////
cv::Mat predicted(trainR.rows, 1, CV_32F); //allocate space for a predicted response
stim::image<unsigned char> test(X, Y, 3); //create a 3-channel (color) image
test = 0;
for(size_t y = 0; y < Y; y++){ //for each pixel in the image
for(size_t x = 0; x < X; x++){
cv::Mat response(3, 1, CV_32FC1);
cv::Mat sample = trainF.row((int)(y * X + x)); //load the corresponding value from the training array
ann->predict(sample, response); //predict the response for this pixel
test(x, y, 0) = (unsigned char)(response.at<float>(0, 0) * 255); //save the response in the color image (make sure to re-scale the values to [0, 255])
test(x, y, 1) = (unsigned char)(response.at<float>(0, 1) * 255);
test(x, y, 2) = (unsigned char)(response.at<float>(0, 2) * 255);
}
}
test.save("response.bmp"); //save the response
}