Face recognition determination threshold

Catalog

face TAR FAR Calculation ：

How to set the threshold in face recognition ？

python draw roc curve ：

---

face TAR FAR Calculation ：

facenet-pytorch-glint360k/test_model_lfw_far_align.py at b1d8b1014b00650688646330fcd258728c7ccb2f · tamerthamoqa/facenet-pytorch-glint360k · GitHub

How to set the threshold in face recognition ？

In face recognition threshold How to set ？

With the wide application of face recognition technology , Most users may not understand one aspect of face recognition , There are also a variety of questions , below , It mainly explains how to set the threshold in visual face recognition ？

First, let's take a look at the set of data provided （ The absolute value has been modified , Qualitatively speaking ）.

False recognition rate	Passing rate	threshold
0.01	0.999	64
0.001	0.998	68
0.0001	0.996	72
1e-5	0.993	76
1e-6	0.985	80
1e-7	0.950	84
1e-8	0.300	88

Through this set of data , The following conclusions can be drawn ：

First, error recognition rate 、 There is a correlation between the passing rate and the threshold , but It belongs to nonlinear relation

Maybe most people see this set of data , The higher the threshold is , The lower the passing rate and error recognition rate , The lower the threshold , The higher the pass rate and error recognition rate .

This leads to a common misunderstanding , The customer thinks that the higher the pass rate , The better the experience , Then the threshold should be set lower . But the higher the passing rate , The higher the error rate , It means that the ability to prevent risks will decline , Therefore, the setting of threshold value should comprehensively consider the customer experience and risk prevention ability . The more accurate statement should be under the specified threshold , company A The pass rate of is higher than that of other companies , The error rate is also lower than that of other companies ,, This means that the company A Our face recognition algorithm is better than other companies .

In addition, the threshold cannot be set too high , From the data, we can see that , After the threshold is continuously raised , The pass rate and the error rate also drop rapidly . At present, face recognition can only be used as an auxiliary authentication means , It cannot be equated with the second generation Key Such a strong authentication tool , If the amount of face data is large enough , There must be misunderstanding , There is no way to risk 0.

The data provided is based on their own face recognition test base , threshold 、 The error rate and pass rate are not absolute . But from a qualitative point of view , Still worthy of reference . If there are a large number of multiple births in the test base , Or the data scale of face database is very large , Then the results will be different .  

python draw roc curve ：

# -*- coding: utf-8 -*-
"""
Created on Thu Sep 21 16:13:04 2017
@author: lizhen
"""
import numpy as np
import matplotlib.pyplot as plt
from sklearn import svm, datasets
from sklearn.metrics import roc_curve, auc  ### Calculation roc and auc
from sklearn import model_selection as cross_validation
# Import some data to play with
iris = datasets.load_iris()
X = iris.data
y = iris.target

## Turn into 2 classification 
X, y = X[y != 2], y[y != 2]

# Add noisy features to make the problem harder
random_state = np.random.RandomState(0)
n_samples, n_features = X.shape
X = np.c_[X, random_state.randn(n_samples, 200 * n_features)]

# shuffle and split training and test sets
X_train, X_test, y_train, y_test = cross_validation.train_test_split(X, y, test_size=.3, random_state=0)

# Learn to predict each class against the other
svm = svm.SVC(kernel='linear', probability=True, random_state=random_state)

### adopt decision_function() Calculated y_score Value , Use in roc_curve() Function 
y_score = svm.fit(X_train, y_train).decision_function(X_test)

# Compute ROC curve and ROC area for each class
fpr, tpr, threshold = roc_curve(y_test, y_score)  ### Calculate the true rate and false positive rate 
roc_auc = auc(fpr, tpr)  ### Calculation auc Value 

plt.figure()
lw = 2
plt.figure(figsize=(10, 10))
plt.plot(fpr, tpr, color='darkorange', lw=lw, label='ROC curve (area = %0.2f)' % roc_auc)  ### The false positive rate is the abscissa , The true rate makes the curve for the ordinate 
plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver operating characteristic example')
plt.legend(loc="lower right")
plt.show()

roc curve ：

from ： use Python draw ROC curve _- Never compromise - The blog of -CSDN Blog _python do roc curve

from sklearn.metrics import roc_curve, auc  #  Data preparation 
import numpy as np
from sklearn import metrics
import matplotlib.pyplot as plt

y = np.array([1, 1,1, 2, 2,2])
scores = np.array([0.1, 0.4,0.5, 0.35, 0.8,0.9])

# roc_curve The input is 
# y:  Sample labels 
# scores:  The probability output of the model to the sample is positive 
# pos_label:  Labels marked as positive examples , In this case, it is marked as 2 Is a positive example 
fpr, tpr, thresholds = metrics.roc_curve(y, scores, pos_label=2)  #  False positive rate 
print(fpr)
# array([0., 0.5, 0.5, 1.]) #  True positive rate 
print(tpr)
# array([0.5, 0.5, 1., 1.]) #  threshold 
print(thresholds)
# array([0.8, 0.4, 0.35, 0.1])
auc = metrics.auc(fpr, tpr)
print(auc)

plt.figure()
lw = 2
plt.plot(fpr, tpr, color='darkorange', lw=lw, label='ROC curve (area = %0.2f)' % auc)
plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver operating characteristic example')
plt.legend(loc="lower right")
plt.show()