def features(self, images):
images_np_array = [image_as_array(image) for image in images]
image_height, image_width = images_np_array[
0].shape[:
2] # All images need to be same size, otherwise dont use batch
batch_face_locations = face_recognition.batch_face_locations(
images_np_array,
batch_size=len(images),
number_of_times_to_upsample=1)
detections = []
for face_locations, image in zip(batch_face_locations,
images_np_array):
if not face_locations:
detections.append([])
continue
crops = self.create_crops(width=image_width,
height=image_height,
face_locations=face_locations)
face_features = self.face_features(image=image,
face_locations=face_locations)
detections.append([
FaceDetection(crop=crop, encoding=features)
for crop, features in zip(crops, face_features)
])
return detections
def print_matches_vid(act_frames,
img_in_frames,
output,
frame_count,
known_face_encodings,
known_face_name,
in_tolerance=float(0.6),
upsample=0):
x = in_tolerance
batch_of_face_locations = face_recognition.batch_face_locations(
img_in_frames, batch_size=2, number_of_times_to_upsample=upsample)
for frame_number_in_batch, face_locations in enumerate(
batch_of_face_locations):
number_of_faces_in_frame = len(face_locations)
frame_number = frame_count - 2 + frame_number_in_batch
print("Found {} face(s) in frame #{}.".format(number_of_faces_in_frame,
frame_number))
pil_image = Image.fromarray(act_frames[frame_number_in_batch])
face_encodings = face_recognition.face_encodings(
img_in_frames[frame_number_in_batch],
known_face_locations=face_locations)
face_names = []
for face_encoding in face_encodings:
matches = face_recognition.compare_faces(known_face_encodings,
face_encoding,
tolerance=x)
name = "Unknown"
face_distances = face_recognition.face_distance(
known_face_encodings, face_encoding)
best_match_index = np.argmin(face_distances)
if matches[best_match_index]:
im = known_face_encodings[best_match_index]
name = known_face_name[str(im)]
face_names.append(name)
for (top, right, bottom, left), name in zip(face_locations,
face_names):
cv2.rectangle(act_frames[frame_number_in_batch], (left, top),
(right, bottom), (0, 0, 255), 2)
cv2.rectangle(act_frames[frame_number_in_batch],
(left, bottom - 25), (right, bottom), (0, 0, 255),
cv2.FILLED)
font = cv2.FONT_HERSHEY_DUPLEX
cv2.putText(act_frames[frame_number_in_batch], name,
(left + 6, bottom - 6), font, 0.5, (255, 255, 255), 1)
print("Writing frame {}".format(frame_number))
output.write(act_frames[frame_number_in_batch])
def __get_location_frames(self):
'''Get the total faces detected and f.indexed locations/original/process based on hits.'''
batch = fr.batch_face_locations(self.__process_frames, 1,
self.__ps.batch_size)
hits = np.nonzero(batch)[0]
locations = np.asarray(batch)[hits]
return (sum(len(x) for x in locations),
zip(locations,
np.asarray(self.__original_frames)[hits],
np.asarray(self.__process_frames)[hits]))
def process_video(vid):
print(f"Processing: {vid}")
basename, _ = os.path.splitext(os.path.basename(vid))
os.makedirs(f"{basename}_faces", exist_ok=True)
if os.path.exists(f"{basename}.maga"):
return
video_capture = cv2.VideoCapture(vid)
faces = 0
frames = []
while video_capture.isOpened():
# Grab a single frame of video
ret, frame_cv = video_capture.read()
# Bail out when the video file ends
if not ret:
break
frame = cv2.cvtColor(frame_cv, cv2.COLOR_BGR2RGB)
frames.append(frame)
print(".", end="", flush=""),
# Tune this to GPU Memory Size.
if len(frames) == BATCH_FRAMES:
print("#", end="", flush="")
batch_of_face_locations = face_recognition.batch_face_locations(
frames, number_of_times_to_upsample=0)
for idx, face_locations in enumerate(batch_of_face_locations):
if len(face_locations) == 0:
continue
face_encodings = face_recognition.face_encodings(
face_image=frames[idx],
known_face_locations=face_locations)
for face_encoding, face_location in zip(
face_encodings, face_locations):
print("$", end="", flush=True)
top, right, bottom, left = face_location
# You can access the actual face itself like this:
face_image = frame[top:bottom, left:right]
pil_image = Image.fromarray(face_image)
pil_image.save(
f"{basename}_faces/{basename}-{faces:08d}.jpg")
with open(f"{basename}_faces/{basename}-{faces:08d}.np",
"wb") as b:
np.save(b,
face_encoding,
allow_pickle=False,
fix_imports=False)
faces += 1
print("", end="\n", flush=True)
frames = []
with open(f"{basename}.maga", "w") as F:
F.write("MAGA!")
def face_recognition(self,
image_list,
using_cuda=True,
batch_size=16,
model="cnn"):
'''
:param image_list: 待处理的图像列表。
:param using_cuda: 是否使用GPU加速，如果使用GPU，则推荐使用batch cnn模型
:param batch_size: 一个batch的大小
:param model: cnn/hog，hog模型在cpu的环境下运行较快
:return:
'''
location_list = []
print("Start Face Detection")
if using_cuda and model == "cnn":
for i in tqdm(range(0, len(image_list), batch_size)):
batch_location = face_recognition.batch_face_locations(
image_list[i:i + batch_size],
number_of_times_to_upsample=0,
batch_size=batch_size)
location_list = location_list + batch_location
else:
for i in tqdm(range(len(image_list))):
image = image_list[i]
location = face_recognition.face_locations(image, model=model)
location_list.append(location)
# face_image, loc = self.face_extract(image, loc)
assert len(image_list) == len(location_list)
speaking_label_list = []
for i in range(len(location_list)):
image, locations = image_list[i], location_list[i]
speaking_label = []
for loc in locations:
face_image, dlib_loc = self.face_extract(image, loc)
check_result = self.face_check(image, dlib_loc)
if not check_result["flag"]:
self.face_save(check_result["embedding"], face_image,
check_result["index"])
abs_label, rel_label = self.speaking_label(
check_result["shape"])
speaking_label.append({
"index": check_result["index"],
"abs_label": abs_label,
"rel_label": rel_label
})
speaking_label_list.append(speaking_label)
return speaking_label_list
def detectSpeakerFace(frame_paths, speaker_enc):
n = len(frame_paths)
frame_batch = []
speaker_bb = {}
for frame_counter in range(n):
frame = fr.load_image_file(frame_paths[frame_counter])
frame_batch.append(frame)
if frame_counter != n - 1 and len(frame_batch) != BATCH_SIZE:
continue
loc_batch = fr.batch_face_locations(
frame_batch, number_of_times_to_upsample=UPSAMPLE)
for frame_number_in_batch, curr_locations in enumerate(loc_batch):
curr_frame_number = frame_counter + 1 - len(
frame_batch) + frame_number_in_batch
curr_frame_path = frame_paths[curr_frame_number]
curr_frame = frame_batch[frame_number_in_batch]
m = ('%-20s %-6d %-3d' %
(curr_frame_path, curr_frame_number, len(curr_locations)))
print(FORMAT % ('detect_frame', m))
if len(curr_locations) == 0:
continue
curr_encodings = fr.face_encodings(
curr_frame, known_face_locations=curr_locations)
res = fr.compare_faces(curr_encodings, speaker_enc)
# TODO: find the res[k] == True such that distance to speaker is minimized
for k in range(len(curr_encodings)):
enc = curr_encodings[k]
loc = curr_locations[k]
curr_frame_name = '/'.join(curr_frame_path.split('/')[-2:])
if res[k] == True:
top, right, bottom, left = curr_locations[k]
speaker_bb[curr_frame_name] = {}
speaker_bb[curr_frame_name]['top'] = top
speaker_bb[curr_frame_name]['right'] = right
speaker_bb[curr_frame_name]['bottom'] = bottom
speaker_bb[curr_frame_name]['left'] = left
frame_batch = []
return speaker_bb
def batch_face_locations(images: List[numpy.ndarray], batch_size=128):
"""Finds all face locations in a list of images through batch processing
Args:
images (List[numpy.ndarray]): List of images
Returns:
List[tuple]: Face boxes (top, right, bottom, left)
"""
if batch_size < 1:
raise ValueError
face_locations = fr.batch_face_locations(images,
number_of_times_to_upsample=0,
batch_size=batch_size)
assert len(face_locations) == len(images)
return face_locations
def getFrameInfo(frame_paths, speaker_enc):
n = len(frame_paths)
frame_batch = []
frame_info = {}
for frame_counter in range(n):
frame = fr.load_image_file(frame_paths[frame_counter])
frame_batch.append(frame)
if frame_counter != n - 1 and len(frame_batch) != BATCH_SIZE:
continue
loc_batch = fr.batch_face_locations(
frame_batch, number_of_times_to_upsample=UPSAMPLE)
for frame_number_in_batch, curr_locations in enumerate(loc_batch):
curr_frame_number = frame_counter + 1 - len(
frame_batch) + frame_number_in_batch
curr_frame_path = frame_paths[curr_frame_number]
curr_frame = frame_batch[frame_number_in_batch]
print("%-20s %-6d %-3d" %
(curr_frame_path, curr_frame_number, len(curr_locations)))
if len(curr_locations) == 0:
continue
curr_encodings = fr.face_encodings(
curr_frame, known_face_locations=curr_locations)
res = fr.compare_faces(curr_encodings, speaker_enc)
for k in range(len(curr_encodings)):
enc = curr_encodings[k]
loc = curr_locations[k]
curr_frame_name = '/'.join(curr_frame_path.split('/')[-2:])
frame_info[curr_frame_name] = {}
if res[k] == True:
top, right, bottom, left = curr_locations[k]
frame_info[curr_frame_name]['top'] = top
frame_info[curr_frame_name]['right'] = right
frame_info[curr_frame_name]['bottom'] = bottom
frame_info[curr_frame_name]['left'] = left
frame_batch = []
return frame_info
def get_face_locations(frames, GPU=False, batch_size=64):
face_coordinates = []
if GPU:
for i in range(0, len(frames), batch_size):
batch_of_frames = frames[i:i+batch_size]
batch_face_locations = face_recognition.batch_face_locations(batch_of_frames, number_of_times_to_upsample=0)
face_coordinates += batch_face_locations
face_coordinates = [f[-1] if f is not None and len(f) else None for f in face_coordinates]
else:
for frame in tqdm(frames):
coordinates_found = face_recognition.face_locations(frame)
if coordinates_found is not None and len(coordinates_found):
face_coordinates.append(coordinates_found[-1])
else:
face_coordinates.append(None)
return face_coordinates
def crop_face_batch(imgs, batch_size):
data_len = imgs.shape[0]
hasFace = np.ones((data_len), dtype=bool)
faces = np.ndarray((data_len, 3, 224, 224), dtype=np.float32)
imgs = numpy_convert_to_list(imgs)
batch_locs = fr.batch_face_locations(imgs, batch_size=batch_size)
for idx, locations in enumerate(batch_locs):
if len(locations) == 0:
hasFace[idx] = False
continue
flags = locations[max_area_indx(locations)]
cropped = cv.resize(imgs[idx][flags[0]:flags[2], flags[3]:flags[1]],
(224, 224)).astype(np.float32)
faces[idx] = np.transpose(cropped, (2, 0, 1))
return faces, hasFace
def processBatch(self, raw_frames, rgb_frames, frame_count, outdir):
target_found = False
batch_of_face_locations = face_recognition.batch_face_locations(
rgb_frames, number_of_times_to_upsample=0)
for frame_number_in_batch, face_locations in enumerate(
batch_of_face_locations):
frame_number = frame_count - len(
rgb_frames) + frame_number_in_batch
raw_frame = raw_frames[frame_number_in_batch]
rgb_frame = rgb_frames[frame_number_in_batch]
outfile = os.path.join(outdir,
"frame_{0}.jpg".format(frame_number))
self.processImage(raw_frame, rgb_frame, outfile)
return target_found
def batch_job(frames, images, positions, face_counts, batch_size):
batch_of_face_locations = face_recognition.batch_face_locations(
frames, number_of_times_to_upsample=0, batch_size=batch_size)
for frame, faces in zip(frames, batch_of_face_locations):
position = []
for (top, right, bottom, left) in faces:
img_face = frame[top:bottom, left:right]
img_yuv = cv2.cvtColor(img_face, cv2.COLOR_BGR2YUV)
# equalize the histogram of the Y channel
img_yuv[:, :, 0] = cv2.equalizeHist(img_yuv[:, :, 0])
# convert the YUV image back to RGB format
img_output = cv2.cvtColor(img_yuv, cv2.COLOR_YUV2BGR)
img_output2 = cv2.resize(img_output, (299, 299),
interpolation=cv2.INTER_AREA)
images.append(img_output2) # faces
position.append((top, right, bottom, left))
face_counts.append(len(faces))
positions.append(position)
def load_face_data():
# 已知人脸库 Start
# 加载一个示例图片并学习如何识别它。
yzc_image = face_recognition.batch_face_locations("/know_face_img/YZC.jpg")
yzc_face_encoding = face_recognition.face_encodings(yzc_image)[0]
# 加载一个示例图片并学习如何识别它。
jack_image = face_recognition.load_image_file("/know_face_img/jack.jpg")
jack_face_encoding = face_recognition.face_encodings(jack_image)[0]
# 加载一个示例图片并学习如何识别它。
wuyj_image = face_recognition.load_image_file(
"/know_face_img/WuYongjun.jpg")
wuyj_face_encoding = face_recognition.face_encodings(wuyj_image)[0]
# 创建已知人脸编码及其名称的数组
known_face_encodings = [
yzc_face_encoding, jack_face_encoding, wuyj_face_encoding
]
known_face_names = ["Youzhengcai", "Jack Ma", "WuYongjun"]
def run():
assert os.path.exists("data")
batch_size = 64
with open("data/celeba_cropped/list_eval_partition_filtered.txt", "r") as f:
lines = f.readlines()
fnames = list([l.split()[0] for l in lines])
print("Generating embeddings...")
embeddings = []
for i in range(0, len(fnames), batch_size):
print(i)
imax = min(i + batch_size, len(fnames))
fname_batch = fnames[i:imax]
images = list(
[face_recognition.load_image_file("data/celeba/img_align_celeba/%s" % fname) for fname in fname_batch]
)
locations = face_recognition.batch_face_locations(images, batch_size=batch_size)
batch_embeddings = []
for j in range(len(fname_batch)):
img = images[j]
loc = locations[j]
embedding_list = face_recognition.face_encodings(img, loc)
if len(embedding_list) == 0:
# Fill with nans
embedding = np.full((128,), np.nan)
else:
embedding = embedding_list[0]
batch_embeddings.append(embedding)
embeddings.append(np.stack(batch_embeddings, axis=0))
embeddings = np.concatenate(embeddings, axis=0)
np.save("data/celeba_cropped/embeddings.npy", embeddings)
with open("data/celeba_cropped/embedding_file_order.txt", "w") as f:
f.write("\n".join(fnames))
def faces_location_vedio(self, vedio_path):
"""每帧一个图片，待调"""
import cv2
video_capture = cv2.VideoCapture(vedio_path)
frames = []
frame_count = 0
while video_capture.isOpened():
# Grab a single frame of video
ret, frame = video_capture.read()
# Bail out when the video file ends
if not ret:
break
# Convert the image from BGR color (which OpenCV uses) to RGB color (which face_recognition uses)
frame = frame[:, :, ::-1]
# Save each frame of the video to a list
frame_count += 1
frames.append(frame)
print(frames)
# Every 128 frames (the default batch size), batch process the list of frames to find faces
if len(frames) == 128:
batch_of_face_locations = face_recognition.batch_face_locations(
frames, number_of_times_to_upsample=0)
# Now let's list all the faces we found in all 128 frames
for frame_number_in_batch, face_locations in enumerate(
batch_of_face_locations):
number_of_faces_in_frame = len(face_locations)
frame_number = frame_count - 128 + frame_number_in_batch
logger.info("I found {} face(s) in frame #{}.".format(
number_of_faces_in_frame, frame_number))
for face_location in face_locations:
# Print the location of each face in this frame
top, right, bottom, left = face_location
logger.info(
" - A face is located at pixel location Top: {}, Left: {}, Bottom: {}, Right: {}"
.format(top, left, bottom, right))
# Clear the frames array to start the next batch
frames = []
def extractFaces(frame_paths):
n = len(frame_paths)
face_encodings = []
enc_to_loc = []
frame_batch = []
for frame_counter in range(n):
frame = fr.load_image_file(frame_paths[frame_counter])
frame_batch.append(frame)
if frame_counter != n - 1 and len(frame_batch) != BATCH_SIZE:
continue
loc_batch = fr.batch_face_locations(frame_batch, number_of_times_to_upsample=UPSAMPLE)
for frame_number_in_batch, curr_locations in enumerate(loc_batch):
curr_frame_number = frame_counter + 1 - len(frame_batch) + frame_number_in_batch
curr_frame_path = frame_paths[curr_frame_number]
curr_frame = frame_batch[frame_number_in_batch]
m = ('%-20s %-6d %-3d' % (curr_frame_path, curr_frame_number, len(curr_locations)))
print(FORMAT % ('proc_frame', m))
if len(curr_locations) == 0:
continue
curr_encodings = fr.face_encodings(curr_frame, known_face_locations=curr_locations)
for k in range(len(curr_encodings)):
enc = curr_encodings[k]
loc = curr_locations[k]
enc_to_loc.append({'frame': curr_frame_number, 'loc': loc})
face_encodings.append(enc)
frame_batch = []
return (face_encodings, enc_to_loc)
def calc_bbox(image_list, batch_size=5):
"""Batch infer of face location, batch_size should be factor of total frame number."""
top_sum = right_sum = bottom_sum = left_sum = 0
for i in tqdm(range(len(image_list) // batch_size)):
image_batch = []
for j in range(i * batch_size, (i + 1) * batch_size):
image = face_recognition.load_image_file(image_list[j])
image_batch.append(image)
face_locations = face_recognition.batch_face_locations(
image_batch, number_of_times_to_upsample=0, batch_size=batch_size)
for face_location in face_locations:
top, right, bottom, left = face_location[
0] # assuming only one face detected in the frame
top_sum += top
right_sum += right
bottom_sum += bottom
left_sum += left
return (top_sum // len(image_list), right_sum // len(image_list),
bottom_sum // len(image_list), left_sum // len(image_list))
img_files = glob("/home/ubuntu/big_data/*.jpg")
dir_name = "/home/ubuntu/face_cropped/"
for start_index in range(0, len(img_files), batch_size):
# get a batch of image filenames from the directory
img_batch = img_files[start_index:start_index + batch_size]
# convert all those image files to numpy arrays
images = {img: face_recognition.load_image_file(img) for img in img_batch}
images = {
fname: images[fname]
for fname in images if images[fname].shape == (1920, 1080, 3)
}
fnames = list(images.keys())
images = list(images.values())
# use face_detection API to get face locations
batch_of_face_locations = face_recognition.batch_face_locations(
images, number_of_times_to_upsample=0, batch_size=len(images))
for i in range(len(batch_of_face_locations)):
face_location = batch_of_face_locations[i][0]
# unpack bounding box coordinates of face_locations
top, right, bottom, left = face_location
# crop the image to only the face
face_image = images[i][top:bottom, left:right]
# generate new filename with original basename in new directory
new_file_name = dir_name + os.path.basename(fnames[i])
# save the image file to disk
img = Image.fromarray(face_image, 'RGB')
img.save(new_file_name)
print("a completed batch at position: " + str(start_index))
while video_capture.isOpened():
# Grab a single frame of video
ret, frame = video_capture.read()
# Bail out when the video file ends
if not ret:
break
# Save each frame of the video to a list
frame_count += 1
frames.append(frame)
# Every 128 frames (the default batch size), batch process the list of frames to find faces
if len(frames) == 128:
batch_of_face_locations = face_recognition.batch_face_locations(frames, number_of_times_to_upsample=0)
# Now let's list all the faces we found in all 128 frames
for frame_number_in_batch, face_locations in enumerate(batch_of_face_locations):
number_of_faces_in_frame = len(face_locations)
frame_number = frame_count - 128 + frame_number_in_batch
print("I found {} face(s) in frame #{}.".format(number_of_faces_in_frame, frame_number))
for face_location in face_locations:
# Print the location of each face in this frame
top, right, bottom, left = face_location
print(" - A face is located at pixel location Top: {}, Left: {}, Bottom: {}, Right: {}".format(top, left, bottom, right))
# Clear the frames array to start the next batch
frames = []
def face_dectect_knn(filename_video_input):
label=0
knn_clf =pickle.load(open(filename_knn_model,'rb'))
knn_clf.n_jobs=16
folder_temp=filename_video_input.split(".")[0]
folder_temp =os.path.join(folder_output,folder_temp)
if not os.path.exists(folder_temp):
os.mkdir(folder_temp)
file_video=os.path.join(folder_videos,filename_video_input)
video = cv2.VideoCapture(file_video)
frames_total=video.get(cv2.CAP_PROP_FRAME_COUNT)
progress=tqdm([i for i in range(int(frames_total))],desc="正在识别视频帧",unit='帧')
frame_count = 1
success = True
begin = time.time()
total={}
frames=[]
while success:
success, frame = video.read()
if (frame_count % frames_every_capture == 0):
capture=int(frame_count/frames_every_capture)
progress.update(frames_every_capture)
frame = cv2.resize(frame, (0, 0),fx=resize_scale, fy=resize_scale,interpolation=cv2.INTER_CUBIC)
# 视频画面可以缩小一些，但是阈值也要调整
frames.append(frame)
if len(frames)>=frames_to_recofnize_once:
batch_of_face_locations = face_recognition.batch_face_locations(frames, number_of_times_to_upsample=0)
frame_index_zip=[]
face_encodings_zip=[]
face_locations_zip=[]
for frame_index, face_locations in enumerate(batch_of_face_locations):
if len(face_locations)==0:
continue
face_encodings = face_recognition.face_encodings(frame, face_locations)
for i in range(len(face_encodings)):
frame_index_zip.append(frame_index)
face_encodings_zip+=face_encodings
face_locations_zip+=face_locations
predicts = knn_clf.predict(face_encodings_zip)
for frame_index,(bottom, right, top, left), result in zip(frame_index_zip,face_locations_zip, predicts):
if result == "ycy":
frame = frames[frame_index]
frame_number = capture - frames_to_recofnize_once + frame_index
face=[bottom,top,left,right]
image_cut = frame[bottom:top, left:right]
label += 1
filename_save = "recognition/true/{}-{}.jpg".format(label,frame_number)
cv2.imwrite(filename_save, image_cut)
toleranses = {
0.4: 10,
}
if frame_number in total:
if "faces" in total[frame_number]:
total[frame_number]["faces"].append(face)
else:
total[frame_number]["faces"]=[face]
else:
total[frame_number]["toleranse"] = toleranses
else:
frame = frames[frame_index]
frame_number = capture - frames_to_recofnize_once + frame_index
image_cut = frame[bottom:top, left:right]
label += 1
filename_save = "recognition/false/{}-{}.jpg".format(label,frame_number)
cv2.imwrite(filename_save, image_cut)
frames = []
frame_count = frame_count + 1
cv2.waitKey(1)
end = time.time()
progress.close()
print("识别视频帧共计用时{}秒".format(round(float(end - begin), 3)))
with io.open(os.path.join(folder_temp,filename_static),"w",encoding="utf-8") as fd:
text = json.dumps(total, ensure_ascii=False, indent=4)
fd.write(text)
video.release()
