1. 모듈 다운로드
!pip install deepface
!pip install dlib고용량 RAM ON
2. 얼굴 랜드마크 시연
랜드마크 모델을 사용해서 이미지에 랜드마크 건설
p = "shape_predictor_68_face_landmarks.dat"
detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor(p)
3. 딥페이스를 이용하여 나이, 성별, 감정 표현 및 민족성 획득
얼굴의 랜드마크를 찾는 모델만있나?
그건 아니다 세상 똑똑이들이 만들어둔 모델들이 많다.
!gdown --id 1Id32-d-nS9BooBLLkw1PQhvLWWAukCsq
!gdown --id 1txWignSWdELl8cWdZHYqIlSE2ZRjI8WI
!gdown --id 1d_tQRWjvQ5i4lZyUfFEfRj7LzXWXseBY
!gdown --id 1kWp2CVg_xTIFqdZAwfN_86A3grim9NyI
!mv facial_expression_model_weights.zip /root/.deepface/weights/facial_expression_model_weights.zip
!mv age_model_weights.h5 /root/.deepface/weights/age_model_weights.h5
!mv gender_model_weights.h5 /root/.deepface/weights/gender_model_weights.h5
!mv race_model_single_batch.zip /root/.deepface/weights/race_model_single_batch.zip그럼 저위의 오리지널 미식가 아저씨의 정보를 예측해보자.
MTCNN 백엔드 사용 얼굴 감지
from deepface import DeepFace
obj = DeepFace.analyze(img_path = "/content/me.jpg", actions = ['age', 'gender', 'race', 'emotion'])
나이..를 제외하고는 다 맞다.
얼추 잘못봤을때 표정이 화난걸로 오해할수도 있는데,
표정의 상태를 "놀람"으로 높은 정확도로 예측했다.
다른 이미지로 예측 및 시각화
민지를 잘생쁨이라고 많이들 말하는데,
모델이 남자 87 , 여자 13로 예측했다. 진짜 맞는것같다.
아시아인, 행복함
Deep Face 백엔드 사용하여 얼굴 감지
달라진게 없다.
똑같이 MTCNN 백엔드를 사용하여 얼굴을 감지한 결과와 똑같다.
4. 유사도 비교
result = DeepFace.verify("/content/me.jpg", "/content/minji.jpg")
print("Is verified: ", result["verified"])
resultVGG-face 모델을 사용해서 미식가 아저씨와 민지의 유사도를 비교한다.
일단 둘이 다르다는것을 알려주고 0.88로 0에서 매우 떨어진 숫자이다.
그렇다면 위의 화이트 민지와 블루 민지를 비교해보자.
result = DeepFace.verify("/content/6.jpg", "/content/minji.jpg")
print("Is verified: ", result["verified"])
result둘이 똑같다고 예측을 하며
0.22로 0과 가까운수치이다.
거리 측정 방식 변경
따로 거리 distance의 형식을 선택해주지 않으면 cosine으로 기본값이 설정된다.
그래서 d_metric인자를 추가해 원하는 것을 넣어주면 된다.
주요 거리 측정 방식
- 코사인 거리 (Cosine Similarity):
일반적으로 텍스트 분석에서 많이 사용되지만, 얼굴 인식에서도 방향성에 중점을 둔 비교에 유용
- 유클리드 거리 (Euclidean Distance):
다양한 분야에서 널리 사용되며, 얼굴 인식에서도 기본적으로 많이 사용
- 유클리드 L2 거리 (Euclidean L2 Distance):
얼굴 임베딩 벡터를 정규화한 후 비교할 때 유용
위 특징을 알고 다음에 상황에 맞게 무언가를 바꿔서 사용할수있다
라는정도로 알아두자.
유사도 한번에 처리
from deepface import DeepFace
import pandas as pd
df = DeepFace.find(img_path = "/content/minji.jpg", db_path = '/content/training_faces', detector_backend = 'ssd')
dftraining_faces 폴더 안의 이미지 중 minji = 화이트 민지와 유사도가 가장 높은 이미지를 찾는 코드
결과
24-06-14 00:47:44 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:47:45 - find function duration 0.09854912757873535 seconds
[ identity hash \
0 /content/training_faces/6.jpg 8356576d1f79488ec043b32531a4c9fbb2df367d
target_x target_y target_w target_h source_x source_y source_w \
0 57 37 72 96 252 113 151
source_h threshold distance
0 220 0.68 0.225453 ]보면 training_faces/6.jpg 가 유사도가 가장 높으며 distance는 0.22이다.
6.jpg는 블루민지이고, 참고로 training_faces 폴더 안에는 다양한 이미지 파일이 있다.
다른 모델 시도
이거는 계속 에러가 발생해 에러코드를 확인하니까
Tensorflow 버전 문제라 버전을 낮춰주고 다시 시작했다.
# TensorFlow 버전 다운그레이드
!pip install tensorflow==2.12
# 런타임 다시 시작 필요
import os
os.kill(os.getpid(), 9)
출력값
24-06-14 00:57:54 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:57:56 - find function duration 1.9901697635650635 seconds
Results for model VGG-Face: [ identity hash \
0 /content/training_faces/6.jpg 8356576d1f79488ec043b32531a4c9fbb2df367d
target_x target_y target_w target_h source_x source_y source_w \
0 57 37 72 96 252 113 151
source_h threshold distance
0 220 0.68 0.225453 ]
24-06-14 00:57:56 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:57:59 - find function duration 2.9760186672210693 seconds
Results for model Facenet: [ identity hash \
0 /content/training_faces/6.jpg 8356576d1f79488ec043b32531a4c9fbb2df367d
target_x target_y target_w target_h source_x source_y source_w \
0 57 37 72 96 252 113 151
source_h threshold distance
0 220 0.4 0.203414 ]
24-06-14 00:57:59 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:58:01 - find function duration 1.4572718143463135 seconds
Results for model OpenFace: [Empty DataFrame
Columns: [identity, hash, target_x, target_y, target_w, target_h, source_x, source_y, source_w, source_h, threshold, distance]
Index: []]
24-06-14 00:58:01 - Found 11 newly added image(s), 0 removed image(s), 0 replaced image(s).
Finding representations: 0%| | 0/11 [00:00<?, ?it/s]24-06-14 00:58:12 - VGGFace2_DeepFace_weights_val-0.9034.h5 will be downloaded...
Downloading...
From: https://github.com/swghosh/DeepFace/releases/download/weights-vggface2-2d-aligned/VGGFace2_DeepFace_weights_val-0.9034.h5.zip
To: /root/.deepface/weights/VGGFace2_DeepFace_weights_val-0.9034.h5.zip
0%| | 0.00/511M [00:00<?, ?B/s]
3%|▎ | 16.3M/511M [00:00<00:03, 161MB/s]
13%|█▎ | 67.1M/511M [00:00<00:01, 362MB/s]
23%|██▎ | 119M/511M [00:00<00:00, 432MB/s]
33%|███▎ | 168M/511M [00:00<00:00, 456MB/s]
44%|████▎ | 223M/511M [00:00<00:00, 488MB/s]
53%|█████▎ | 273M/511M [00:00<00:00, 490MB/s]
63%|██████▎ | 322M/511M [00:00<00:00, 490MB/s]
73%|███████▎ | 374M/511M [00:00<00:00, 497MB/s]
83%|████████▎ | 425M/511M [00:00<00:00, 501MB/s]
100%|██████████| 511M/511M [00:01<00:00, 473MB/s]
Finding representations: 18%|█▊ | 2/11 [00:22<01:24, 9.38s/it]24-06-14 00:58:23 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_3.jpg: Face could not be detected in /content/training_faces/Nidia_3.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 64%|██████▎ | 7/11 [00:28<00:09, 2.41s/it]24-06-14 00:58:29 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_2.jpg: Face could not be detected in /content/training_faces/Nidia_2.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 100%|██████████| 11/11 [00:34<00:00, 3.12s/it]
24-06-14 00:58:35 - There are now 11 representations in ds_model_deepface_detector_ssd_aligned_normalization_base_expand_0.pkl
24-06-14 00:58:35 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:58:37 - find function duration 35.9246301651001 seconds
Results for model DeepFace: [ identity hash \
0 /content/training_faces/6.jpg 8356576d1f79488ec043b32531a4c9fbb2df367d
target_x target_y target_w target_h source_x source_y source_w \
0 57 37 72 96 252 113 151
source_h threshold distance
0 220 0.23 0.202124 ]
24-06-14 00:58:37 - Found 11 newly added image(s), 0 removed image(s), 0 replaced image(s).
Finding representations: 0%| | 0/11 [00:00<?, ?it/s]24-06-14 00:58:37 - deepid_keras_weights.h5 will be downloaded...
Downloading...
From: https://github.com/serengil/deepface_models/releases/download/v1.0/deepid_keras_weights.h5
To: /root/.deepface/weights/deepid_keras_weights.h5
100%|██████████| 1.61M/1.61M [00:00<00:00, 39.9MB/s]
Finding representations: 36%|███▋ | 4/11 [00:02<00:02, 2.52it/s]24-06-14 00:58:39 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_3.jpg: Face could not be detected in /content/training_faces/Nidia_3.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 73%|███████▎ | 8/11 [00:02<00:00, 5.53it/s]24-06-14 00:58:39 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_2.jpg: Face could not be detected in /content/training_faces/Nidia_2.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 100%|██████████| 11/11 [00:03<00:00, 3.42it/s]
24-06-14 00:58:40 - There are now 11 representations in ds_model_deepid_detector_ssd_aligned_normalization_base_expand_0.pkl
24-06-14 00:58:40 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:58:40 - find function duration 3.3007586002349854 seconds
Results for model DeepID: [Empty DataFrame
Columns: [identity, hash, target_x, target_y, target_w, target_h, source_x, source_y, source_w, source_h, threshold, distance]
Index: []]
24-06-14 00:58:40 - Found 11 newly added image(s), 0 removed image(s), 0 replaced image(s).
Finding representations: 0%| | 0/11 [00:00<?, ?it/s]24-06-14 00:58:41 - arcface_weights.h5 will be downloaded to /root/.deepface/weights/arcface_weights.h5
Downloading...
From: https://github.com/serengil/deepface_models/releases/download/v1.0/arcface_weights.h5
To: /root/.deepface/weights/arcface_weights.h5
0%| | 0.00/137M [00:00<?, ?B/s]
18%|█▊ | 24.1M/137M [00:00<00:00, 240MB/s]
39%|███▉ | 53.5M/137M [00:00<00:00, 269MB/s]
62%|██████▏ | 85.5M/137M [00:00<00:00, 292MB/s]
100%|██████████| 137M/137M [00:00<00:00, 295MB/s]
Finding representations: 18%|█▊ | 2/11 [00:04<00:15, 1.68s/it]24-06-14 00:58:44 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_3.jpg: Face could not be detected in /content/training_faces/Nidia_3.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 82%|████████▏ | 9/11 [00:05<00:00, 3.76it/s]24-06-14 00:58:45 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_2.jpg: Face could not be detected in /content/training_faces/Nidia_2.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 100%|██████████| 11/11 [00:06<00:00, 1.72it/s]
24-06-14 00:58:46 - There are now 11 representations in ds_model_arcface_detector_ssd_aligned_normalization_base_expand_0.pkl
24-06-14 00:58:46 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:58:46 - find function duration 6.631281137466431 seconds
Results for model ArcFace: [ identity hash \
0 /content/training_faces/6.jpg 8356576d1f79488ec043b32531a4c9fbb2df367d
target_x target_y target_w target_h source_x source_y source_w \
0 57 37 72 96 252 113 151
source_h threshold distance
0 220 0.68 0.270221 ]
24-06-14 00:58:46 - Found 11 newly added image(s), 0 removed image(s), 0 replaced image(s).
Finding representations: 0%| | 0/11 [00:00<?, ?it/s]24-06-14 00:58:47 - dlib_face_recognition_resnet_model_v1.dat is going to be downloaded
Downloading...
From: http://dlib.net/files/dlib_face_recognition_resnet_model_v1.dat.bz2
To: /root/.deepface/weights/dlib_face_recognition_resnet_model_v1.dat.bz2
0%| | 0.00/21.4M [00:00<?, ?B/s]
2%|▏ | 524k/21.4M [00:00<00:34, 599kB/s]
5%|▍ | 1.05M/21.4M [00:01<00:17, 1.13MB/s]
10%|▉ | 2.10M/21.4M [00:01<00:08, 2.30MB/s]
17%|█▋ | 3.67M/21.4M [00:01<00:04, 4.02MB/s]
34%|███▍ | 7.34M/21.4M [00:01<00:01, 8.55MB/s]
51%|█████▏ | 11.0M/21.4M [00:01<00:00, 12.0MB/s]
69%|██████▊ | 14.7M/21.4M [00:01<00:00, 14.5MB/s]
100%|██████████| 21.4M/21.4M [00:02<00:00, 10.0MB/s]
Finding representations: 36%|███▋ | 4/11 [00:06<00:08, 1.16s/it]24-06-14 00:58:52 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_3.jpg: Face could not be detected in /content/training_faces/Nidia_3.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 73%|███████▎ | 8/11 [00:06<00:01, 2.23it/s]24-06-14 00:58:53 - 🔴 Exception while extracting faces from /content/training_faces/Nidia_2.jpg: Face could not be detected in /content/training_faces/Nidia_2.jpg.Please confirm that the picture is a face photo or consider to set enforce_detection param to False.
Finding representations: 100%|██████████| 11/11 [00:07<00:00, 1.51it/s]
24-06-14 00:58:54 - There are now 11 representations in ds_model_dlib_detector_ssd_aligned_normalization_base_expand_0.pkl
24-06-14 00:58:54 - Searching /content/minji.jpg in 11 length datastore
24-06-14 00:58:54 - find function duration 7.35061502456665 seconds
Results for model Dlib: [ identity \
0 /content/training_faces/6.jpg
1 /content/training_faces/Nidia_4.jpg
2 /content/training_faces/Nidia_1.jpg
hash target_x target_y target_w \
0 8356576d1f79488ec043b32531a4c9fbb2df367d 57 37 72
1 992c672cbb40152de9c37d6628d14daf080a1728 321 74 167
2 8c5eaabf7e11dfb6fc7ffaaacc233ab4d0c0f43e 136 91 63
target_h source_x source_y source_w source_h threshold distance
0 96 252 113 151 220 0.07 0.030508
1 246 252 113 151 220 0.07 0.062589
2 105 252 113 151 220 0.07 0.065299 ]
identity \
0 /content/training_faces/6.jpg
1 /content/training_faces/6.jpg
2 /content/training_faces/6.jpg
3 /content/training_faces/6.jpg
4 /content/training_faces/6.jpg
5 /content/training_faces/Nidia_4.jpg
6 /content/training_faces/Nidia_1.jpg
hash target_x target_y target_w \
0 8356576d1f79488ec043b32531a4c9fbb2df367d 57 37 72
1 8356576d1f79488ec043b32531a4c9fbb2df367d 57 37 72
2 8356576d1f79488ec043b32531a4c9fbb2df367d 57 37 72
3 8356576d1f79488ec043b32531a4c9fbb2df367d 57 37 72
4 8356576d1f79488ec043b32531a4c9fbb2df367d 57 37 72
5 992c672cbb40152de9c37d6628d14daf080a1728 321 74 167
6 8c5eaabf7e11dfb6fc7ffaaacc233ab4d0c0f43e 136 91 63
target_h source_x source_y source_w source_h threshold distance \
0 96 252 113 151 220 0.68 0.225453
1 96 252 113 151 220 0.40 0.203414
2 96 252 113 151 220 0.23 0.202124
3 96 252 113 151 220 0.68 0.270221
4 96 252 113 151 220 0.07 0.030508
5 246 252 113 151 220 0.07 0.062589
6 105 252 113 151 220 0.07 0.065299
model
0 VGG-Face
1 Facenet
2 DeepFace
3 ArcFace
4 Dlib
5 Dlib
6 Dlib 요약하자면 VGG-Face, Facenet, DeepFace, ArcFace, Dlib 모델을 사용하여 /content/minji.jpg 이미지를 /content/training_faces 폴더 내의 이미지들과 비교
대부분의 모델(VGG-Face, Fac
enet, DeepFace, ArcFace, Dlib)은 블루민지와 화이트 민지가 동일 인물임을 인식했다.
OpenFace와 DeepID 모델은 유사한 얼굴을 찾지 못했다.
Dlib 모델은 /content/training_faces/Nidia_4.jpg와 /content/training_faces/Nidia_1.jpg에서도 낮은 거리 값을 반환했지만, 임계값 기준으로는 동일 인물로 인식하지 않았다.
이 결과를 바탕으로 각 모델의 성능을 평가할 수 있으며, 특정 모델이 더 잘 맞는지 확인할 수 있다.
감사합니다. https://www.youtube.com/channel/UCxlkiu9_aWijoD7BannNM7w