Spaces:
Sleeping
Sleeping
Create deepface_template_utils.py
Browse files- deepface_template_utils.py +82 -0
deepface_template_utils.py
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import json
|
| 3 |
+
import numpy as np
|
| 4 |
+
from deepface import DeepFace
|
| 5 |
+
import cv2
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
def auto_rotate_if_needed(image_path):
|
| 9 |
+
img = cv2.imread(image_path)
|
| 10 |
+
if img is None:
|
| 11 |
+
raise ValueError(f"Image non lisible : {image_path}")
|
| 12 |
+
|
| 13 |
+
h, w = img.shape[:2]
|
| 14 |
+
|
| 15 |
+
# 1. Si image couchée → pivoter verticalement
|
| 16 |
+
if w > h:
|
| 17 |
+
img = cv2.rotate(img, cv2.ROTATE_90_CLOCKWISE)
|
| 18 |
+
|
| 19 |
+
# 2. Si image à l’envers (tête en bas), corriger :
|
| 20 |
+
# On peut ici forcer une rotation supplémentaire si nécessaire
|
| 21 |
+
# 💡 OPTION : ajouter une vérification automatique plus intelligente avec un modèle plus tard
|
| 22 |
+
img = cv2.rotate(img, cv2.ROTATE_180) # Pour le cas à l’envers
|
| 23 |
+
|
| 24 |
+
# 3. Écraser l'image modifiée
|
| 25 |
+
cv2.imwrite(image_path, img)
|
| 26 |
+
return image_path
|
| 27 |
+
|
| 28 |
+
def auto_rotate_folder(folder_path):
|
| 29 |
+
for filename in os.listdir(folder_path):
|
| 30 |
+
if filename.lower().endswith((".jpg", ".jpeg", ".png")):
|
| 31 |
+
full_path = os.path.join(folder_path, filename)
|
| 32 |
+
auto_rotate_if_needed(full_path)
|
| 33 |
+
|
| 34 |
+
def auto_rotate_all(folders):
|
| 35 |
+
"""
|
| 36 |
+
Corrige l'orientation des images dans une liste de dossiers,
|
| 37 |
+
y compris leurs sous-dossiers.
|
| 38 |
+
"""
|
| 39 |
+
for folder_path in folders:
|
| 40 |
+
print(f"\n🔁 Correction des images dans : {folder_path}")
|
| 41 |
+
for root, _, files in os.walk(folder_path):
|
| 42 |
+
for filename in files:
|
| 43 |
+
if filename.lower().endswith((".jpg", ".jpeg", ".png")):
|
| 44 |
+
full_path = os.path.join(root, filename)
|
| 45 |
+
try:
|
| 46 |
+
auto_rotate_if_needed(full_path)
|
| 47 |
+
except Exception as e:
|
| 48 |
+
print(f"⚠️ Erreur sur {filename} : {e}")
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def create_template_from_folder(folder_path, model_name="ArcFace"):
|
| 54 |
+
embeddings = []
|
| 55 |
+
for filename in os.listdir(folder_path):
|
| 56 |
+
if filename.lower().endswith((".jpg", ".jpeg", ".png")):
|
| 57 |
+
img_path = os.path.join(folder_path, filename)
|
| 58 |
+
try:
|
| 59 |
+
print(f"📷 Traitement : {filename}")
|
| 60 |
+
auto_rotate_if_needed(img_path)
|
| 61 |
+
reps = DeepFace.represent(img_path=img_path, model_name=model_name, enforce_detection=True)
|
| 62 |
+
embeddings.append(reps[0]["embedding"])
|
| 63 |
+
except Exception as e:
|
| 64 |
+
print(f"⚠️ Erreur sur {filename} : {e}")
|
| 65 |
+
if not embeddings:
|
| 66 |
+
raise RuntimeError("❌ Aucun embedding valide extrait.")
|
| 67 |
+
return np.mean(np.array(embeddings), axis=0)
|
| 68 |
+
|
| 69 |
+
def save_template(template_array, output_path):
|
| 70 |
+
with open(output_path, "w") as f:
|
| 71 |
+
json.dump(template_array.tolist(), f)
|
| 72 |
+
print(f"✅ Template sauvegardé dans : {output_path}")
|
| 73 |
+
|
| 74 |
+
def load_template(input_path):
|
| 75 |
+
with open(input_path, "r") as f:
|
| 76 |
+
return np.array(json.load(f))
|
| 77 |
+
|
| 78 |
+
def verify_against_template(image_path, template_array, threshold=0.5, model_name="ArcFace"):
|
| 79 |
+
reps = DeepFace.represent(img_path=image_path, model_name=model_name, enforce_detection=True)
|
| 80 |
+
embedding = np.array(reps[0]["embedding"])
|
| 81 |
+
similarity = np.dot(template_array, embedding) / (np.linalg.norm(template_array) * np.linalg.norm(embedding))
|
| 82 |
+
return similarity >= threshold, similarity
|