目录
一、说明
本文主要对基于C++和Python的虹膜识别的效果进行测试,并对比其结果。基于C++的虹膜识别测试工程为:基于C++和Opencv2的虹膜识别工程。基于Python的测试代码为:https://github.com/thuyngch/Iris-Recognition。
二、测试处理
为了使用相同的数据集进行测试,本文以Python版本的虹膜识别为主,保存测试过程中使用的测试图片,然后用C++版本对这些图片进行测试,测试结果的评价标准也按照Python版本中的标准:
fscore = 2*precision*recall / (precision+recall)
1、基于Python的虹膜识别
主要以代码https://github.com/thuyngch/Iris-Recognition为主,使用说明在原代码中已经很详细,本文使用的环境为:Window10、Python3.7.3。需要注意的几点如下:
(1)虹膜识别的测试文件为:eval_casia1.py,需要到python文件加下运行:python3 eval_casia1.py。注意代码中测试数据集CASIA-IrisV1存放路径:
CASIA1_DIR = "../CASIA1"
(2)直接运行python3 eval_casia1.py会出错,根据错误提示,添加了
if __name__ == '__main__':
即代码为:
if __name__ == '__main__':
#------------------------------------------------------------------------------
# Main execution
#------------------------------------------------------------------------------
# Get identities of MMU2 dataset
identities = glob(os.path.join(CASIA1_DIR, "**"))
identities = sorted([os.path.basename(identity) for identity in identities])
n_identities = len(identities)
print("Number of identities:", n_identities)
# Construct a dictionary of files
files_dict = {}
image_files = []
for identity in identities:
files = glob(os.path.join(CASIA1_DIR, identity, "*.*"))
shuffle(files)
files_dict[identity] = files[:N_IMAGES]
image_files += files[:N_IMAGES]
n_image_files = len(image_files)
print("Number of image files:", n_image_files)
# save test images to file 'pro_imgs.txt'
txt_open = open('./pro_imgs.txt', 'w')
for i in image_files:
line = i + '\n'
txt_open.write(line)
txt_open.close()
# Ground truth
ground_truth = np.zeros([n_image_files, n_image_files], dtype=int)
for i in range(ground_truth.shape[0]):
for j in range(ground_truth.shape[1]):
if i//N_IMAGES == j//N_IMAGES:
ground_truth[i, j] = 1
# Evaluate parameters
pools = Pool(processes=cpu_count())
best_results = []
for eye_threshold in tqdm(eyelashes_thresholds, total=len(eyelashes_thresholds)):
# Extract features
args = zip(image_files, repeat(eye_threshold), repeat(False))
features = list(pools.map(pool_func_extract_feature, args))
# Calculate the distances
args = []
for i in range(n_image_files):
for j in range(n_image_files):
if i>=j:
continue
arg = (features[i][0], features[i][1], features[j][0], features[j][1])
args.append(arg)
distances = pools.map(pool_func_calHammingDist, args)
# Construct a distance matrix
k = 0
dist_mat = np.zeros([n_image_files, n_image_files])
for i in range(n_image_files):
for j in range(n_image_files):
if i<j:
dist_mat[i, j] = distances[k]
k += 1
elif i>j:
dist_mat[i, j] = dist_mat[j, i]
# Metrics
accuracies, precisions, recalls, fscores = [], [], [], []
for threshold in thresholds:
decision_map = (dist_mat<=threshold).astype(int)
accuracy = (decision_map==ground_truth).sum() / ground_truth.size
precision = (ground_truth*decision_map).sum() / decision_map.sum()
recall = (ground_truth*decision_map).sum() / ground_truth.sum()
fscore = 2*precision*recall / (precision+recall)
accuracies.append(accuracy)
precisions.append(precision)
recalls.append(recall)
fscores.append(fscore)
# Save the best result
best_fscore = max(fscores)
best_threshold = thresholds[fscores.index(best_fscore)]
best_accuracy = accuracies[fscores.index(best_fscore)]
best_precision = precisions[fscores.index(best_fscore)]
best_recall = recalls[fscores.index(best_fscore)]
best_results.append((eye_threshold, best_threshold, best_fscore, best_accuracy, best_precision, best_recall))
# Show the final best result
eye_thresholds = [item[0] for item in best_results]
thresholds = [item[1] for item in best_results]
fscores = [item[2] for item in best_results]
accuracies = [item[3] for item in best_results]
precisions = [item[4] for item in best_results]
recalls = [item[5] for item in best_results]
print("Maximum fscore: ", max(fscores))
print("Best accuracy: ", accuracies[fscores.index(max(fscores))])
print("Best precision: ", precisions[fscores.index(max(fscores))])
print("Best recall: ", recalls[fscores.index(max(fscores))])
print("Best eye_threshold: ", eye_thresholds[fscores.index(max(fscores))])
print("Best threshold: ", thresholds[fscores.index(max(fscores))])
CASIA-IrisV1数据集共有108人,每人为7张虹膜图片,代码中随机从7张图片中选取4个:
N_IMAGES = 4
故总测试图片数为:
108*4=432
将随机选取的图片进行保存:
# save test images to file 'pro_imgs.txt'
txt_open = open('./pro_imgs.txt', 'w')
for i in image_files:
line = i + '\n'
txt_open.write(line)
txt_open.close()
代码中有两个参数:eyelashes_thresholds和thresholds。eyelashes_thresholds为眼睫毛的阈值,范围为[10,250],选取25个数值:
eyelashes_thresholds = np.linspace(start=10, stop=250, num=25)
[ 10. 20. 30. 40. 50. 60. 70. 80. 90. 100. 110. 120. 130. 140. 150. 160. 170. 180. 190. 200. 210. 220. 230. 240. 250.]
thresholds为判断两个虹膜是否为同一个的阈值,两个虹膜的差异值小于thresholds则为同一人,否则为不同人。取值范围为[0, 1],选取100个,步长为:
(1-0)/(100-1)= 0.01010101
thresholds = np.linspace(start=0.0, stop=1.0, num=100)
[0. 0.01010101 0.02020202 0.03030303 0.04040404 0.05050505 0.06060606 0.07070707 0.08080808 0.09090909 0.1010101 0.11111111 0.12121212 0.13131313 0.14141414 0.15151515 0.16161616 0.17171717 0.18181818 0.19191919 0.2020202 0.21212121 0.22222222 0.23232323 0.24242424 0.25252525 0.26262626 0.27272727 0.28282828 0.29292929 0.3030303 0.31313131 0.32323232 0.33333333 0.34343434 0.35353535 0.36363636 0.37373737 0.38383838 0.39393939 0.4040404 0.41414141 0.42424242 0.43434343 0.44444444 0.45454545 0.46464646 0.47474747 0.48484848 0.49494949 0.50505051 0.51515152 0.52525253 0.53535354 0.54545455 0.55555556 0.56565657 0.57575758 0.58585859 0.5959596 0.60606061 0.61616162 0.62626263 0.63636364 0.64646465 0.65656566 0.66666667 0.67676768 0.68686869 0.6969697 0.70707071 0.71717172 0.72727273 0.73737374 0.74747475 0.75757576 0.76767677 0.77777778 0.78787879 0.7979798 0.80808081 0.81818182 0.82828283 0.83838384 0.84848485 0.85858586 0.86868687 0.87878788 0.88888889 0.8989899 0.90909091 0.91919192 0.92929293 0.93939394 0.94949495 0.95959596 0.96969697 0.97979798 0.98989899 1. ]
(3)注意运行路径
2、基于C++虹膜识别
具体代码详见:基于C++和Opencv2的虹膜识别工程。原代码中测试图片为:下一个与上一个进行比较。为了和Python代码保持一致,需要做下修改。
(1)将CASIA1测试数据图片复制到data文件夹中;
(2)修改process.ini文件。修改图片存放文件名称为Python中保存的文件:
Load List of images = pro_imgs.txt
修改图片存放路径:
Load original images = CASIA1/
屏蔽所有保存路径:
#Save segmented images = Output/SegmentedImages/
#Save contours parameters = Output/CircleParameters/
#Save masks of iris = Output/Masks/
#Save normalized images = Output/NormalizedImages/
#Save normalized masks = Output/NormalizedMasks/
#Save iris codes = Output/IrisCodes/
#Save matching scores = Output/score.txt
(3)添加my_run()。在OsiManager.h中加入void my_run();如下图所示
在OsiManager.cpp中添加:
void OsiManager::my_run()
{
cout << endl;
cout << "================" << endl;
cout << "Start processing" << endl;
cout << "================" << endl;
cout << endl;
// If matching is requested, create a file
ofstream result_matching;
if (mProcessMatching && mOutputFileMatchingScores != "")
{
try
{
result_matching.open(mOutputFileMatchingScores.c_str(), ios::out);
}
catch (exception & e)
{
cout << e.what() << endl;
throw runtime_error("Cannot create the file for matching scores : " + mOutputFileMatchingScores);
}
}
int num_test_imgs = mListOfImages.size();
int N_IMAGES = 4;
float threshold = 0.393939393939394;
int decision_map_num = 0;
int ground_truth_num = 0;
// ground_truth and dist_mat initial
bool(*ground_truth)[432] = new bool[num_test_imgs][432];
float(*dist_mat)[432] = new float[num_test_imgs][432];
bool(*decision_map)[432] = new bool[num_test_imgs][432];
for (int i = 0; i<num_test_imgs; ++i)
{
for (int j = 0; j<num_test_imgs; ++j)
{
if (i / N_IMAGES == j / N_IMAGES)
{
ground_truth[i][j] = 1;
ground_truth_num++;
}
else
ground_truth[i][j] = 0;
dist_mat[i][j] = 0;
decision_map[i][j] = 0;
}
}
//end ground_truth and dist_mat initial
vector<OsiEye> eyr_res;
cout << "Extract Features start!" << endl;
for (int i = 0; i < num_test_imgs; i++)
{
OsiEye eye;
processOneEye(mListOfImages[i], eye);
eyr_res.push_back(eye);
cout << i + 1 << " pic is over!!!" << endl;
}
cout << "Extract Features end!!!" << endl;
for (int i = 0; i < num_test_imgs; i++)
{
for (int j = 0; j < num_test_imgs; ++j)
{
cout << "start: " << i << " " << j << " ";
if (i < j)
{
OsiEye eyr_i = eyr_res[i];
OsiEye eyr_j = eyr_res[j];
float score = (eyr_res[i]).match((eyr_res[j]), mpApplicationPoints);
dist_mat[i][j] = score;
}
else if (i > j)
{
dist_mat[i][j] = dist_mat[j][i];
}
if (dist_mat[i][j] < threshold)
{
decision_map[i][j] = 1;
decision_map_num++;
}
cout << dist_mat[i][j] << " " << decision_map[i][j] << endl;
}
}
// result
int accuracy_num = 0;
int precision_num = 0;
for (int i = 0; i < num_test_imgs; i++)
{
for (int j = 0; j < num_test_imgs; ++j)
{
if (decision_map[i][j] == ground_truth[i][j])
{
accuracy_num++;
}
if (decision_map[i][j] && ground_truth[i][j])
{
precision_num++;
}
}
}
float accuracy = float(accuracy_num) / float(num_test_imgs*num_test_imgs);
float precision_res = float(precision_num) / float(decision_map_num);
float recall = float(precision_num) / float(ground_truth_num);
float fscore = 2 * precision_res*recall / (precision_res + recall);
cout << "Result: " << endl << "accuracy: " << accuracy << "\t"
<< "precision_res: " << precision_res << "\t"
<< "recall: " << recall << "\t"
<< "fscore: " << fscore << endl;
// end result
//将阈值从0到1按照0.01步长递增,寻找最优的fscore值,并获取对应的最优阈值
float fscore_best = 0.0;
float threshold_best = 0.0;
float recall_best = 0.0;
float precision_best = 0.0;
float accuracy_best = 0.0;
for (threshold = 0.0; threshold < 1.0; threshold += 0.01010101)
{
int accuracy_num = 0;
int precision_num = 0;
decision_map_num = 0;
for (int i = 0; i < num_test_imgs; ++i)
{
for (int j = 0; j < num_test_imgs; ++j)
{
if (dist_mat[i][j] < threshold)
{
decision_map[i][j] = 1;
decision_map_num++;
}
if (decision_map[i][j] == ground_truth[i][j])
{
accuracy_num++;
}
if (decision_map[i][j] && ground_truth[i][j])
{
precision_num++;
}
}
}
float accuracy = float(accuracy_num) / float(num_test_imgs*num_test_imgs);
float precision_res = float(precision_num) / float(decision_map_num);
float recall = float(precision_num) / float(ground_truth_num);
float fscore = 2 * precision_res*recall / (precision_res + recall);
cout << "Result: " << "accuracy: " << accuracy << "\t"
<< "precision_res: " << precision_res << "\t"
<< "recall: " << recall << "\t"
<< "fscore: " << fscore << endl;
if (fscore > fscore_best)
{
fscore_best = fscore;
threshold_best = threshold;
recall_best = recall;
precision_best = precision_res;
accuracy_best = accuracy;
}
for (int i = 0; i<num_test_imgs; ++i)
{
for (int j = 0; j<num_test_imgs; ++j)
{
decision_map[i][j] = 0;
}
}
}
cout << "Best fscore is: " << fscore_best << ", Best threshold is: " << threshold_best <<
", Best recall is: " << recall_best << ", Best precision is: " << precision_best << ", Best accuracy is: " << accuracy_best << endl;
// save dist_mat
ofstream result_dist;
string result_dist_path = "../data/dist_mat.txt";
try
{
result_dist.open(result_dist_path.c_str(), ios::out);
}
catch (exception & e)
{
cout << e.what() << endl;
throw runtime_error("Cannot create the file for result_dist_path scores : " + result_dist_path);
}
if (result_dist)
{
try
{
for (int i = 0; i < num_test_imgs; i++)
{
int j = 0;
for (j = 0; j < num_test_imgs - 1; ++j)
{
result_dist << dist_mat[i][j] << " ";
}
result_dist << dist_mat[i][j] << "\n";
}
}
catch (exception & e)
{
cout << e.what() << endl;
throw runtime_error("Error while saving result of matching in " + mOutputFileMatchingScores);
}
}
// end save dist_mat
cout << endl;
cout << "==============" << endl;
cout << "End processing" << endl;
cout << "==============" << endl;
cout << endl;
} // end of function
注意:这里只选取了Python代码中的thresholds作为最优结果选择参数,其他参数使用原C++代码中默认的。
(4)若在上述代码中保存了decision_map.txt文件,则寻找最优阈值的代码如下:
void test_best_thr(string test_resuilt_file)
{
// Open the file
ifstream file(test_resuilt_file.c_str(), ifstream::in);
if (!file.good())
throw runtime_error("Cannot read configuration file " + test_resuilt_file);
int decision_map_num = 0;
int ground_truth_num = 0;
int N_IMAGES = 4;
int num_test_imgs = 432; //测试图片总数,定义选取N_IMAGES时,总数为:N_IMAGES*108(108为数据集CASIA1中人的数量,即文件夹的数量)
// ground_truth and dist_mat initial
bool(*ground_truth)[432] = new bool[num_test_imgs][432]; //定义ground_truth变量
float(*dist_mat)[432] = new float[num_test_imgs][432]; //定义测试结果dist_mat变量
bool(*decision_map)[432] = new bool[num_test_imgs][432]; // 根据测试结果和设置的阈值得到的正确与否(0/1)的变量
// 变量初始化
for (int i = 0; i<num_test_imgs; ++i)
{
for (int j = 0; j<num_test_imgs; ++j)
{
if (i / N_IMAGES == j / N_IMAGES)
{
ground_truth[i][j] = 1;
ground_truth_num++;
}
else
ground_truth[i][j] = 0;
dist_mat[i][j] = 0;
decision_map[i][j] = 0;
}
}
// 读取结果文件并按照结果文件保存的格式重新将结果取到dist_mat变量中
int x = 0;
bool flag = true;
while (file.good() && !file.eof() && flag)
{
// Get the new line
string line;
getline(file, line);
// Filter out comments
if (!line.empty())
{
int y = 0;
size_t start = 0, index = line.find_first_of(" ", 0);
while (index != line.npos)
{
if (start != index)
{
std::string s = line.substr(start, index - start);
float res = atof(s.c_str());
dist_mat[x][y] = res;
y++;
start = index + 1;
index = line.find_first_of(" ", start);
}
else
break;
}
if (!line.substr(start).empty())
{
std::string s = line.substr(start);
float res = atof(s.c_str());
dist_mat[x][y] = res;
y++;
}
if (y != num_test_imgs)
{
cout << y << " != " << num_test_imgs << endl;
flag = false;
break;
}
}
x++;
}
if (x != num_test_imgs)
{
cout << x << " != " << num_test_imgs << endl;
return;
}
//将阈值从0到1按照0.01步长递增,寻找最优的fscore值,并获取对应的最优阈值
float fscore_best = 0.0;
float threshold_best = 0.0;
float recall_best = 0.0;
float precision_best = 0.0;
float accuracy_best = 0.0;
for (float threshold = 0; threshold < 1.0; threshold += 0.01010101)
{
int accuracy_num = 0;
int precision_num = 0;
decision_map_num = 0;
for (int i = 0.0; i < num_test_imgs; ++i)
{
for (int j = 0; j < num_test_imgs; ++j)
{
if (dist_mat[i][j] < threshold)
{
decision_map[i][j] = 1;
decision_map_num++;
}
if (decision_map[i][j] == ground_truth[i][j])
{
accuracy_num++;
}
if (decision_map[i][j] && ground_truth[i][j])
{
precision_num++;
}
}
}
float accuracy = float(accuracy_num) / float(num_test_imgs*num_test_imgs);
float precision_res = float(precision_num) / float(decision_map_num);
float recall = float(precision_num) / float(ground_truth_num);
float fscore = 2 * precision_res*recall / (precision_res + recall);
cout << "Result: " << "accuracy: " << accuracy << "\t"
<< "precision_res: " << precision_res << "\t"
<< "recall: " << recall << "\t"
<< "fscore: " << fscore << endl;
if (fscore > fscore_best)
{
fscore_best = fscore;
threshold_best = threshold;
recall_best = recall;
precision_best = precision_res;
accuracy_best = accuracy;
}
for (int i = 0; i<num_test_imgs; ++i)
{
for (int j = 0; j<num_test_imgs; ++j)
{
decision_map[i][j] = 0;
}
}
}
cout << "Best fscore is: " << fscore_best << ", Best threshold is: " << threshold_best <<
", Best recall is: " << recall_best << ", Best precision is: " << precision_best << ", Best accuracy is: " << accuracy_best << endl;
}
(5)main中调用
将osi.run()改为osi.my_run();
三、测试结果
测试结果如下表所示,在CASIA-IrisV1上,C++版本的效果要优于Python版本,具体原因,后续看代码比较再来说明。