Upload 2 files

api.py

@@ -49,6 +49,33 @@ class JobStatus(str, Enum):
 
 jobs: Dict[str, dict] = {}
 
+def normalize_face_lighting(image):
+    """
+    Normaliza el brillo de una imagen de cara usando CLAHE.
+    Esto reduce el impacto de diferentes condiciones de iluminación en los embeddings.
+    
+    Args:
+        image: Imagen BGR (OpenCV format)
+    
+    Returns:
+        Imagen normalizada en el mismo formato
+    """
+    import cv2
+    # Convertir a LAB color space (más robusto para iluminación)
+    lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
+    l, a, b = cv2.split(lab)
+    
+    # Aplicar CLAHE (Contrast Limited Adaptive Histogram Equalization) al canal L
+    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
+    l_normalized = clahe.apply(l)
+    
+    # Recombinar canales
+    lab_normalized = cv2.merge([l_normalized, a, b])
+    
+    # Convertir de vuelta a BGR
+    normalized = cv2.cvtColor(lab_normalized, cv2.COLOR_LAB2BGR)
+    return normalized
+
 def hierarchical_cluster_with_min_size(X, max_groups: int, min_cluster_size: int):
     """
     Clustering jerárquico aglomerativo que produce hasta max_groups clusters.
@@ -231,6 +258,7 @@ def process_video_job(job_id: str):
         # Detección de caras y embeddings (CPU), alineado con 'originales'
         try:
             print(f"[{job_id}] Iniciando detección de personajes (CPU, originales)...")
+            print(f"[{job_id}] Normalización de brillo activada (CLAHE) para reducir impacto de iluminación")
             import cv2
             import numpy as np
             try:
@@ -274,7 +302,9 @@ def process_video_job(job_id: str):
                 if not ret2:
                     continue
                 frames_processed += 1
-                rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+                # Normalizar iluminación antes de procesar
+                frame_normalized = normalize_face_lighting(frame)
+                rgb = cv2.cvtColor(frame_normalized, cv2.COLOR_BGR2RGB)
 
                 if _use_fr and face_recognition is not None:
                     boxes = face_recognition.face_locations(rgb, model="hog")  # CPU HOG
@@ -283,7 +313,7 @@ def process_video_job(job_id: str):
                         frames_with_faces += 1
                         print(f"[{job_id}] Frame {frame_idx}: {len(boxes)} cara(s) detectada(s) con face_recognition")
                     for (top, right, bottom, left), e in zip(boxes, encs):
-                        crop = frame[top:bottom, left:right]
+                        crop = frame_normalized[top:bottom, left:right]
                         if crop.size == 0:
                             continue
                         fn = f"face_{frame_idx:06d}_{saved_count:03d}.jpg"
@@ -304,7 +334,7 @@ def process_video_job(job_id: str):
                         pass
                     else:
                         try:
-                            gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
+                            gray = cv2.cvtColor(frame_normalized, cv2.COLOR_BGR2GRAY)
                             try:
                                 haar_path = getattr(cv2.data, 'haarcascades', None) or ''
                                 face_cascade = cv2.CascadeClassifier(os.path.join(haar_path, 'haarcascade_frontalface_default.xml'))
@@ -322,7 +352,7 @@ def process_video_job(job_id: str):
                             if not boxes_haar:
                                 try:
                                     tmp_detect = faces_root / f"detect_{frame_idx:06d}.jpg"
-                                    cv2.imwrite(str(tmp_detect), frame)
+                                    cv2.imwrite(str(tmp_detect), frame_normalized)
                                     detect_result = DeepFace.extract_faces(img_path=str(tmp_detect), detector_backend='opencv', enforce_detection=False)
                                     for det in detect_result:
                                         facial_area = det.get('facial_area', {})
@@ -344,7 +374,7 @@ def process_video_job(job_id: str):
                                 print(f"[{job_id}] Frame {frame_idx}: {len(boxes_haar)} cara(s) detectada(s) con Haar/DeepFace")
                             
                             for (top, right, bottom, left) in boxes_haar:
-                                crop = frame[top:bottom, left:right]
+                                crop = frame_normalized[top:bottom, left:right]
                                 if crop.size == 0:
                                     continue
                                 fn = f"face_{frame_idx:06d}_{saved_count:03d}.jpg"