def threshold2(img_dir):
t1 = time.time()
mtcnn = load_mtcnn()
t2 = time.time()
print('mtcnn loaded, time: %.2f' % (t2 - t1))
imgs = load_imgs(mtcnn, img_dir, False)
t1 = time.time()
print('images loaded, count: %d, time: %.2f' % (len(imgs), t1 - t2))
with tf.Graph().as_default():
with tf.Session() as sess:
t1 = time.time()
model = load_pretrain_model()
t2 = time.time()
print('pretrained model loaded, time: %.2f' % (t2 - t1))
embeddings = calc_embeddings(model, sess, imgs)
diff = np.subtract(embeddings[0::2], embeddings[1::2])
dist = np.sum(np.square(diff), 1)
print('dist same person', dist)
em1 = [embeddings[0], embeddings[1]]
em2 = [embeddings[2], embeddings[3]]
diff = np.subtract(em1, em2)
dist = np.sum(np.square(diff), 1)
print('dist diff person', dist)
def threshold(img_dir):
t1 = time.time()
mtcnn = load_mtcnn()
t2 = time.time()
print('mtcnn loaded, time: %.2f' % (t2 - t1))
imgs = load_imgs(mtcnn, img_dir)
random.shuffle(imgs)
t1 = time.time()
print('images loaded, count: %d, time: %.2f' % (len(imgs), t1 - t2))
batch_count = 10
batch_size = math.floor(len(imgs) / 10)
thresholds = np.arange(0.3, 1.2, 0.01)
threshold_accuracies = []
with tf.Graph().as_default():
with tf.Session() as sess:
t1 = time.time()
model = load_pretrain_model()
t2 = time.time()
print('pretrained model loaded, time: %.2f' % (t2 - t1))
embeddings = calc_embeddings(model, sess, imgs)
for i in range(batch_count - 1):
t1 = time.time()
ta = find_best_threshold(
imgs[batch_size * i:batch_size * (i + 1)],
embeddings[batch_size * i:batch_size * (i + 1)],
thresholds)
t2 = time.time()
threshold_accuracies.append(ta)
print('batch: %d, threshold %f, accuracy: %f, time: %.2f' %
(i, ta[0], ta[1], t2 - t1))
ta = np.mean(threshold_accuracies, 0)
threshold = ta[0]
accuracy = ta[1]
print('mean threshold: %f, mean accuracy: %f' %
(threshold, accuracy))
#test
accuracy = calc_accuracy(imgs[batch_size * 9:],
embeddings[batch_size * 9:], threshold)
print('test accuracy: ', accuracy)
def to_lfw(src_dir, dst_dir):
if (not os.path.isdir(dst_dir)):
os.makedirs(dst_dir)
mtcnn = load_mtcnn()
dirs = os.listdir(src_dir)
for person in dirs:
person_dir = os.path.join(src_dir, person)
if (not os.path.isdir(person_dir)):
continue
dst_person_dir = os.path.join(dst_dir, person)
if (not os.path.isdir(dst_person_dir)):
os.mkdir(dst_person_dir)
files = os.listdir(person_dir)
for file in files:
img_file = os.path.join(person_dir, file)
detect_align(mtcnn, img_file, dst_person_dir)
def main():
REDIS_EXPIRE = 10
mtcnn = load_mtcnn()
r = redis.Redis('localhost', 6379, password='******')
while (True):
print('wait to detect face!')
filename = r.brpop('face_detect')
filename = bytes.decode(filename[1])
# print('filename', filename)
imgpath = os.path.join('../uploads', filename)
# print('imgpath', imgpath)
start = time.time()
faces = detect(mtcnn, imgpath, '../uploads/detect')
print('detect result: ', faces)
count = faces['count']
if count < 0:
ret = {'status': -4, 'msg': 'detect face error'}
else:
ret = {
'status': 0,
'msg': 'success',
'faces': count,
'time': time.time() - start,
'detectUrl': '/detect/' + filename,
'rects': []
}
if (count > 0):
ret['size'] = faces['size']
for _, v in enumerate(faces['rects']):
ret['rects'].append({
'left': int(v[0]),
'top': int(v[1]),
'right': int(v[2]),
'bottom': int(v[3])
})
r.expire(filename, REDIS_EXPIRE)
print(ret)
r.lpush(filename, json.dumps(ret))
def calc_embeddings(model, sess, imgs, image_size=160):
images_placeholder = model[0]
embeddings = model[1]
phase_train_placeholder = model[2]
nrof_images = len(imgs)
images = np.zeros((nrof_images, image_size, image_size, 3))
for i in range(nrof_images):
images[i, :, :, :] = imgs[i]
feed_dict = {images_placeholder: images, phase_train_placeholder: False}
emb_array = sess.run(embeddings, feed_dict=feed_dict)
return emb_array
def calc_distance(embeddings1, embeddings2):
diff = np.subtract(embeddings1, embeddings2)
return np.sum(np.square(diff), 1)
if __name__ == '__main__':
mtcnn = load_mtcnn()
img1 = load_img(mtcnn, '../face_test_img/bbh/1.jpg')
img2 = load_img(mtcnn, '../face_test_img/bbh/2.jpg')
with tf.Graph().as_default():
with tf.Session() as sess:
model = load_pretrain_model()
embeddings = calc_embeddings(model, sess, [img1, img2])
dist = calc_distance([embeddings[0]], [embeddings[1]])
print('distance', dist[0])
def main():
REDIS_EXPIRE = 10
mtcnn = load_mtcnn()
with tf.Graph().as_default():
with tf.Session() as sess:
model = load_pretrain_model()
r = redis.Redis('localhost', 6379, password='******')
while (True):
print('wait to compare face!')
filename = r.brpop('face_compare')
orig_filename = bytes.decode(filename[1])
print('origfilename', orig_filename)
filename = orig_filename.split('#')
filename1 = filename[0]
filename2 = filename[1]
threshold = float(filename[2])
# print('split', filename, filename1, filename2)
imgpath1 = os.path.join('../uploads', filename1)
imgpath2 = os.path.join('../uploads', filename2)
# print('imgpath', imgpath1, imgpath2)
start = time.time()
rect1 = {'left': 0, 'top': 0, 'right': 0, 'bottom': 0}
size1 = {'width': 0, 'height': 0}
img1 = detect_align(mtcnn, imgpath1, '../uploads/align', False,
160, True, rect1, size1)
rect2 = {'left': 0, 'top': 0, 'right': 0, 'bottom': 0}
size2 = {'width': 0, 'height': 0}
img2 = detect_align(mtcnn, imgpath2, '../uploads/align', False,
160, True, rect2, size2)
if (img1 is None or img2 is None):
ret = {'status': -4, 'msg': 'detect face error'}
else:
embeddings = calc_embeddings(model, sess, [img1, img2])
dist = calc_distance([embeddings[0]], [embeddings[1]])[0]
judge_time = time.time() - start
# print('judge time: ', judge_time)
if (dist <= threshold):
msg = "同一人可能性高"
# elif dist <= 1.1:
# msg = "同一人可能性一般"
# elif dist <= 1.23:
# msg = "不同人可能性一般"
else:
msg = "不同人可能性高"
ret = {
'status': 0,
'msg': msg,
'distance': dist.item(),
'time': judge_time,
'size1': size1,
'rect1': rect1,
'size2': size2,
'rect2': rect2
}
r.expire(orig_filename, REDIS_EXPIRE)
print(ret)
r.lpush(orig_filename, json.dumps(ret))