def make_one_to_more_test(self, max_size):
self._load_dataset_more(max_size)
for i, (picture, name) in enumerate(self.dataset_more_base):
# print(picture)
try:
self.face_recognizer.fingerprints[
name + "?" + str(i)] = self._get_face_feature(picture)
except:
pass
# print("ok")
correct_num = 0
cnt = 0
for picture, name in self.dataset_more_test:
# print(f"{cnt} / {len(self.dataset_more_test)}")
cnt += 1
feature = self._get_face_feature(picture)
similarity_max = 0
name_of_max = ""
for p, vector in self.face_recognizer.fingerprints.items():
similarity = FaceRecognizer._cosine_similarity(feature, vector)
if similarity > similarity_max:
name_of_max = p
similarity_max = similarity
# print(f"{similarity_max}, {nam e_of_max}, {p}")
if name_of_max.startswith(name):
correct_num += 1
return correct_num / len(self.dataset_more_test)
class LFWTester:
def __init__(self):
self._load_dataset_index()
self.face_recognizer = FaceRecognizer(test_model_weight_path)
def _load_dataset_index(self) -> None:
# 加载1:1测试数据
dataset_pairs = test_dataset_path + "pairs.txt"
with open(dataset_pairs, "r") as f:
lines = f.read().splitlines()
line_num = 1
self.dataset_one = []
for _ in range(10):
fold = []
for _ in range(300):
name, pic_num1, pic_num2 = lines[line_num].split("\t")
line_num += 1
face1_path = f"{test_dataset_path}{name}/{name}_{pic_num1.zfill(4)}.png"
face2_path = f"{test_dataset_path}{name}/{name}_{pic_num2.zfill(4)}.png"
if os.path.exists(face1_path) and os.path.exists(
face2_path):
fold.append((face1_path, face2_path, True))
for _ in range(300):
name1, pic_num1, name2, pic_num2 = lines[line_num].split(
"\t")
line_num += 1
face1_path = f"{test_dataset_path}{name1}/{name1}_{pic_num1.zfill(4)}.png"
face2_path = f"{test_dataset_path}{name2}/{name2}_{pic_num2.zfill(4)}.png"
if os.path.exists(face1_path) and os.path.exists(
face2_path):
fold.append((face1_path, face2_path, False))
self.dataset_one.append(fold)
def _load_dataset_more(self, max_size):
# 加载1:N测试数据
self.dataset_more_base, self.dataset_more_test = [], []
for cnt, name in enumerate(os.listdir(test_dataset_path)):
if os.path.isfile(test_dataset_path + name):
continue
if cnt > max_size:
break
pictures = os.listdir(test_dataset_path + name)
pictures = [
picture for picture in pictures if picture.endswith(".png")
]
for i, picture in enumerate(sorted(pictures)):
face_path = f"{test_dataset_path}{name}/{picture}"
if os.path.isfile(face_path):
if i == 0 and len(pictures) > 1:
self.dataset_more_test.append([face_path, name])
else:
self.dataset_more_base.append([face_path, name])
def load_dataset_mask(self, test_num=2000):
random.seed(444)
images = dataset.get_mask_slim_datalist()
data = dict()
for image_path, label in images:
if label in data:
data[label].append(image_path)
else:
data[label] = [image_path]
self.dataset_one = []
idx = list(data.keys())
for i in range(int(test_num / 2)):
p1 = random.choice(idx)
p2 = p1
while p2 == p1:
p2 = random.choice(idx)
self.dataset_one.append(
(random.choice(data[p1]), random.choice(data[p2]), False))
for i in range(int(test_num / 2)):
p1 = random.choice(idx)
while len(data[p1]) == 1:
p1 = random.choice(idx)
i1 = random.choice(data[p1])
i2 = i1
while i2 == i1:
i2 = random.choice(data[p1])
self.dataset_one.append((i1, i2, True))
def _one_to_one_test(self, test_data):
result = {}
# 计算相似度
similarities = []
for face1_path, face2_path, label in test_data:
similarity = self._get_similarity(face1_path, face2_path)
similarities.append(similarity)
# 计算准确率
accuracy_rate, similarity_threshold = self._calculate_max_accuracy(
similarities, (data[2] for data in test_data))
# FRR @ FAR = x
return accuracy_rate, similarity_threshold, similarities
def make_one_to_one_test(self):
result = []
for i in range(10):
result.append(self._one_to_one_test(self.dataset_one[i]))
# 计算准确率
test_num = 0
correct_num = 0
for i in range(10):
similarity_threshold = result[i][1]
for j in range(10):
if i != j:
for similarity, (_, _,
label) in zip(result[j][2],
self.dataset_one[j]):
if (similarity >= similarity_threshold) == label:
correct_num += 1
test_num += 1
correct_rate = correct_num / test_num
FRR_FAR_curve = []
# 计算FRR(拒识率)@FAR(误识率)=x
for threshold in np.arange(0, 1, 0.01):
false_acceptance_num = 0
false_rejection_num = 0
true_acceptance_num = 0
true_rejection_num = 0
for i in range(10):
for similarity, (_, _, label) in zip(result[i][2],
self.dataset_one[i]):
if label == True:
if similarity >= threshold:
true_acceptance_num += 1
else:
false_acceptance_num += 1
else:
if similarity >= threshold:
false_rejection_num += 1
else:
true_rejection_num += 1
FRR = false_rejection_num / (false_rejection_num +
true_rejection_num)
FAR = false_acceptance_num / (false_acceptance_num +
true_acceptance_num)
FRR_FAR_curve.append((FAR, FRR))
return correct_rate, np.array(FRR_FAR_curve)
def make_one_to_more_test(self, max_size):
self._load_dataset_more(max_size)
for i, (picture, name) in enumerate(self.dataset_more_base):
# print(picture)
try:
self.face_recognizer.fingerprints[
name + "?" + str(i)] = self._get_face_feature(picture)
except:
pass
# print("ok")
correct_num = 0
cnt = 0
for picture, name in self.dataset_more_test:
# print(f"{cnt} / {len(self.dataset_more_test)}")
cnt += 1
feature = self._get_face_feature(picture)
similarity_max = 0
name_of_max = ""
for p, vector in self.face_recognizer.fingerprints.items():
similarity = FaceRecognizer._cosine_similarity(feature, vector)
if similarity > similarity_max:
name_of_max = p
similarity_max = similarity
# print(f"{similarity_max}, {nam e_of_max}, {p}")
if name_of_max.startswith(name):
correct_num += 1
return correct_num / len(self.dataset_more_test)
@staticmethod
def show_FRR_FAR_curve(FRR_FAR_curve: np.ndarray) -> None:
plt.plot(FRR_FAR_curve[:, 0], FRR_FAR_curve[:, 1])
plt.xlabel("FAR")
plt.ylabel("FRR")
plt.show()
@staticmethod
def _calculate_max_accuracy(similarities, labels):
similarities_with_labels = sorted(list(zip(similarities, labels)),
key=lambda x: x[0])
correct_num = sum(x[1] for x in similarities_with_labels)
max_correct_num = correct_num
similarity_threshold = 0
for similarity, label in similarities_with_labels:
if label == False:
correct_num += 1
else:
correct_num -= 1
if correct_num > max_correct_num:
max_correct_num = correct_num
similarity_threshold = similarity
accuracy_rate = max_correct_num / len(similarities_with_labels)
return accuracy_rate, similarity_threshold
def _get_face_feature(self, image_path):
if image_path in features:
return features[image_path]
image = cv2.imread(image_path)
feature = self.face_recognizer._get_feature_vector(image)
features[image_path] = feature
return feature
def _get_similarity(self, face1_path, face2_path):
face_pair = (face1_path, face2_path)
if face_pair in similarities:
return similarities[face_pair]
feature1 = self._get_face_feature(face1_path)
feature2 = self._get_face_feature(face2_path)
similarity = self.face_recognizer._cosine_similarity(
feature1, feature2)
similarities[face_pair] = similarity
return similarity
