在图像/视频中裁剪出人脸地区

1. 在图像中裁剪人脸地区

1. import face_alignment
2. import skimage.io
3. import numpy
4. from argparse import ArgumentParser
5. from skimage import img_as_ubyte
6. from skimage.transform import resize
7. from tqdm import tqdm
8. import os
9. import numpy as np
10. import warnings
11. warnings.filterwarnings("ignore")
13. # =================================================================================================
14. # Kun(20250306): 输出裁剪图像的命令
15. # =================================================================================================
17. def extract_bbox(frame, fa):
18.     if max(frame.shape[0], frame.shape[1]) > 640:
19.         scale_factor =  max(frame.shape[0], frame.shape[1]) / 640.0
20.         frame = resize(frame, (int(frame.shape[0] / scale_factor), int(frame.shape[1] / scale_factor)))
21.         frame = img_as_ubyte(frame)
22.     else:
23.         scale_factor = 1
24.     frame = frame[..., :3]
25.     bboxes = fa.face_detector.detect_from_image(frame[..., ::-1])
26.     if len(bboxes) == 0:
27.         return []
28.     return np.array(bboxes)[:, :-1] * scale_factor
30. def bb_intersection_over_union(boxA, boxB):
31.     xA = max(boxA[0], boxB[0])
32.     yA = max(boxA[1], boxB[1])
33.     xB = min(boxA[2], boxB[2])
34.     yB = min(boxA[3], boxB[3])
35.     interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)
36.     boxAArea = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)
37.     boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)
38.     iou = interArea / float(boxAArea + boxBArea - interArea)
39.     return iou
41. def join(tube_bbox, bbox):
42.     xA = min(tube_bbox[0], bbox[0])
43.     yA = min(tube_bbox[1], bbox[1])
44.     xB = max(tube_bbox[2], bbox[2])
45.     yB = max(tube_bbox[3], bbox[3])
46.     return (xA, yA, xB, yB)
48. def compute_bbox(tube_bbox, frame_shape, inp, image_shape, increase_area=0.1):
49.     left, top, right, bot = tube_bbox
50.     width = right - left
51.     height = bot - top
53.     #Computing aspect preserving bbox
54.     width_increase = max(increase_area, ((1 + 2 * increase_area) * height - width) / (2 * width))
55.     height_increase = max(increase_area, ((1 + 2 * increase_area) * width - height) / (2 * height))
57.     left = int(left - width_increase * width)
58.     top = int(top - height_increase * height)
59.     right = int(right + width_increase * width)
60.     bot = int(bot + height_increase * height)
62.     top, bot, left, right = max(0, top), min(bot, frame_shape[0]), max(0, left), min(right, frame_shape[1])
63.     h, w = bot - top, right - left
65.     scale = f'{image_shape[0]}:{image_shape[1]}'
67.     return f'ffmpeg -i {inp} -filter:v "crop={w}:{h}:{left}:{top}, scale={scale}" crop.png'
69. def process_image(args):
70.     device = 'cpu' if args.cpu else 'cuda'
71.     fa = face_alignment.FaceAlignment(face_alignment.LandmarksType.TWO_D, flip_input=False, device=device)
72.     frame = skimage.io.imread(args.inp)
73.     frame_shape = frame.shape
75.     bboxes = extract_bbox(frame, fa)
76.     if len(bboxes) == 0:
77.         print("No faces detected.")
78.         return []
80.     tube_bbox = bboxes[0]  # Assuming we take the first detected face
81.     command = compute_bbox(tube_bbox, frame_shape, inp=args.inp, image_shape=args.image_shape, increase_area=args.increase)
82.     return [command]
84. if __name__ == "__main__":
85.     parser = ArgumentParser()
87.     parser.add_argument("--image_shape", default=(256, 256), type=lambda x: tuple(map(int, x.split(','))),
88.                         help="Image shape")
89.     parser.add_argument("--increase", default=0.1, type=float, help='Increase bbox by this amount')
90.     parser.add_argument("--iou_with_initial", type=float, default=0.25, help="The minimal allowed iou with inital bbox")
91.     parser.add_argument("--inp", required=True, help='Input image')
92.     parser.add_argument("--cpu", dest="cpu", action="store_true", help="cpu mode.")
94.     args = parser.parse_args()
96.     commands = process_image(args)
97.     for command in commands:
98.         print(command)

复制代码

2. 在视频中裁剪人脸地区

1. import face_alignment
2. import skimage.io
3. import numpy
4. from argparse import ArgumentParser
5. from skimage import img_as_ubyte
6. from skimage.transform import resize
7. from tqdm import tqdm
8. import os
9. import imageio
10. import numpy as np
11. import warnings
12. warnings.filterwarnings("ignore")
14. # =================================================================================================
15. # Kun(20250306): 输出裁剪视频的命令
16. # =================================================================================================
18. def extract_bbox(frame, fa):
19.     if max(frame.shape[0], frame.shape[1]) > 640:
20.         scale_factor =  max(frame.shape[0], frame.shape[1]) / 640.0
21.         frame = resize(frame, (int(frame.shape[0] / scale_factor), int(frame.shape[1] / scale_factor)))
22.         frame = img_as_ubyte(frame)
23.     else:
24.         scale_factor = 1
25.     frame = frame[..., :3]
26.     bboxes = fa.face_detector.detect_from_image(frame[..., ::-1])
27.     if len(bboxes) == 0:
28.         return []
29.     return np.array(bboxes)[:, :-1] * scale_factor
31. def bb_intersection_over_union(boxA, boxB):
32.     xA = max(boxA[0], boxB[0])
33.     yA = max(boxA[1], boxB[1])
34.     xB = min(boxA[2], boxB[2])
35.     yB = min(boxA[3], boxB[3])
36.     interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)
37.     boxAArea = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)
38.     boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)
39.     iou = interArea / float(boxAArea + boxBArea - interArea)
40.     return iou
42. def join(tube_bbox, bbox):
43.     xA = min(tube_bbox[0], bbox[0])
44.     yA = min(tube_bbox[1], bbox[1])
45.     xB = max(tube_bbox[2], bbox[2])
46.     yB = max(tube_bbox[3], bbox[3])
47.     return (xA, yA, xB, yB)
49. def compute_bbox(start, end, fps, tube_bbox, frame_shape, inp, image_shape, increase_area=0.1):
50.     left, top, right, bot = tube_bbox
51.     width = right - left
52.     height = bot - top
54.     #Computing aspect preserving bbox
55.     width_increase = max(increase_area, ((1 + 2 * increase_area) * height - width) / (2 * width))
56.     height_increase = max(increase_area, ((1 + 2 * increase_area) * width - height) / (2 * height))
58.     left = int(left - width_increase * width)
59.     top = int(top - height_increase * height)
60.     right = int(right + width_increase * width)
61.     bot = int(bot + height_increase * height)
63.     top, bot, left, right = max(0, top), min(bot, frame_shape[0]), max(0, left), min(right, frame_shape[1])
64.     h, w = bot - top, right - left
66.     start = start / fps
67.     end = end / fps
68.     time = end - start
70.     scale = f'{image_shape[0]}:{image_shape[1]}'
72.     return f'ffmpeg -i {inp} -ss {start} -t {time} -filter:v "crop={w}:{h}:{left}:{top}, scale={scale}" crop.mp4'
74. def compute_bbox_trajectories(trajectories, fps, frame_shape, args):
75.     commands = []
76.     for i, (bbox, tube_bbox, start, end) in enumerate(trajectories):
77.         if (end - start) > args.min_frames:
78.             command = compute_bbox(start, end, fps, tube_bbox, frame_shape, inp=args.inp, image_shape=args.image_shape, increase_area=args.increase)
79.             commands.append(command)
80.     return commands
82. def process_video(args):
83.     device = 'cpu' if args.cpu else 'cuda'
84.     fa = face_alignment.FaceAlignment(face_alignment.LandmarksType.TWO_D, flip_input=False, device=device)
85.     video = imageio.get_reader(args.inp)
87.     trajectories = []
88.     previous_frame = None
89.     fps = video.get_meta_data()['fps']
90.     commands = []
91.     try:
92.         for i, frame in tqdm(enumerate(video)):
93.             frame_shape = frame.shape
94.             bboxes =  extract_bbox(frame, fa)
95.             ## For each trajectory check the criterion
96.             not_valid_trajectories = []
97.             valid_trajectories = []
99.             for trajectory in trajectories:
100.                 tube_bbox = trajectory[0]
101.                 intersection = 0
102.                 for bbox in bboxes:
103.                     intersection = max(intersection, bb_intersection_over_union(tube_bbox, bbox))
104.                 if intersection > args.iou_with_initial:
105.                     valid_trajectories.append(trajectory)
106.                 else:
107.                     not_valid_trajectories.append(trajectory)
109.             commands += compute_bbox_trajectories(not_valid_trajectories, fps, frame_shape, args)
110.             trajectories = valid_trajectories
112.             ## Assign bbox to trajectories, create new trajectories
113.             for bbox in bboxes:
114.                 intersection = 0
115.                 current_trajectory = None
116.                 for trajectory in trajectories:
117.                     tube_bbox = trajectory[0]
118.                     current_intersection = bb_intersection_over_union(tube_bbox, bbox)
119.                     if intersection < current_intersection and current_intersection > args.iou_with_initial:
120.                         intersection = bb_intersection_over_union(tube_bbox, bbox)
121.                         current_trajectory = trajectory
123.                 ## Create new trajectory
124.                 if current_trajectory is None:
125.                     trajectories.append([bbox, bbox, i, i])
126.                 else:
127.                     current_trajectory[3] = i
128.                     current_trajectory[1] = join(current_trajectory[1], bbox)
130.     except IndexError as e:
131.         raise (e)
133.     commands += compute_bbox_trajectories(trajectories, fps, frame_shape, args)
134.     return commands
136. if __name__ == "__main__":
137.     parser = ArgumentParser()
139.     parser.add_argument("--image_shape", default=(256, 256), type=lambda x: tuple(map(int, x.split(','))),
140.                         help="Image shape")
141.     parser.add_argument("--increase", default=0.1, type=float, help='Increase bbox by this amount')
142.     parser.add_argument("--iou_with_initial", type=float, default=0.25, help="The minimal allowed iou with inital bbox")
143.     parser.add_argument("--inp", required=True, help='Input image or video')
144.     parser.add_argument("--min_frames", type=int, default=150,  help='Minimum number of frames')
145.     parser.add_argument("--cpu", dest="cpu", action="store_true", help="cpu mode.")
147.     args = parser.parse_args()
149.     commands = process_video(args)
150.     for command in commands:
151.         print (command)
153.         

复制代码