""" Config - Face Recognition System (Revisi) ===================================== Quality-Aware Score-Level Fusion Architecture Perubahan utama: 1. Hapus Dense Layer - langsung normalize dan fuse 2. Score-Level Fusion - similarity, bukan feature concatenation 3. Quality-Aware Weights: wCNN = QualityScore, wLBPH = 1 - QualityScore 4. Quality sederhana: Blur, Brightness, Pose (3 metrik) 5. Multi-frame: 15 → 5 terbaik 6. Hapus dynamic threshold adjustment 7. Incremental update: similarity ≥ 0.98 8. Gallery Quality: simpan quality score, tanggal, jumlah registrasi 9. Top-3 (bukan Top-5) 10. MobileFaceNet → ArcFace Loss (penjelasan) Author: Skripsi 2026 """ from pathlib import Path import numpy as np # === PATH CONFIGURATION === _CURRENT_DIR = Path(__file__).resolve().parent MODEL_DIR = _CURRENT_DIR FACE_MAHASISWA_DIR = MODEL_DIR / "face_mahasiswa" # === MODEL PATHS === YUNET_MODEL_PATH = MODEL_DIR / "models" / "face_detection_yunet_2023mar.onnx" MOBILEFACENET_MODEL_PATH = MODEL_DIR / "models" / "mobilefacenet.onnx" # === IMAGE SIZES === IMG_SIZE_LBPH = 96 # Grayscale input for LBPH IMG_SIZE_CNN = 112 # RGB input for MobileFaceNet IMG_SIZE_ALIGN = 112 # For alignment reference # === LBPH PARAMETERS === LBPH_RADIUS = 1 LBPH_NEIGHBORS = 8 LBPH_GRID_X = 8 LBPH_GRID_Y = 8 LBPH_DIM = (LBPH_NEIGHBORS + 2) * LBPH_GRID_X * LBPH_GRID_Y # 640 # === MOBILEFACENET PARAMETERS === # MobileFaceNet menggunakan ArcFace Loss untuk training # Output: 512-D face embedding yang discriminative CNN_DIM = 512 # ============================================================================= # 1. FACE ALIGNMENT CONFIG (YuNet 5 Landmarks) # ============================================================================= ALIGNMENT_CONFIG = { 'enabled': True, 'target_size': IMG_SIZE_ALIGN, # Reference points untuk frontal face (112x112 template) # STANDARD InsightFace / ArcFace — WAJIB SAMA dengan training distribution # Sumber: insightface/utils/face_align.py arcface_dst # 5 landmark: left_eye, right_eye, nose, left_mouth, right_mouth 'reference_points': np.array([ # Official InsightFace/ArcFace 112x112 alignment template. [38.2946, 51.6963], # left_eye [73.5318, 51.5014], # right_eye [56.0252, 71.7366], # nose [41.5493, 92.3655], # left_mouth [70.7299, 92.2041], # right_mouth ], dtype=np.float64), 'reference_scale': 112.0, } # ============================================================================= # 2. FACE QUALITY ASSESSMENT (Sederhana - 3 Metrik) # ============================================================================= # Revisi: Hanya 3 metrik, semakin sedikit semakin mudah divalidasi QUALITY_CONFIG = { 'enabled': True, # Blur Detection (Variance of Laplacian) # < 30 = blur (reject) 'blur': { 'threshold': 30.0, }, # Brightness Assessment # Revisi Low-Light: diturunkan dari min=40 → 20, max=220 → 230 # Alasan: Di gedung jurusan indoor, mean brightness wajah bisa 20-45px. # Threshold lama (40) terlalu ketat dan menolak wajah yang sebenarnya valid. 'brightness': { 'min': 20, # Revisi: 40 → 20 (toleransi kondisi indoor gelap) 'max': 230, # Revisi: 220 → 230 (toleransi sedikit overexposed) }, # Pose Estimation (dari landmark offset) # Yaw/Pitch > 15° = reject 'pose': { 'yaw_threshold': 15.0, 'pitch_threshold': 15.0, }, # Pass threshold 'pass_threshold': 0.5, # Minimal 2 dari 3 metrik harus bagus } # ============================================================================= # 3. SCORE-LEVEL FUSION CONFIG (BUKAN Feature-Level) # ============================================================================= # Revisi: Score-Level Fusion, bukan Feature-Level Fusion # # Arsitektur lama (Feature-Level): # LBPH → Dense → BN → ReLU ─┐ # ├─► Concat ─► L2Norm ─► Embedding # CNN → Dense → BN → ReLU ─┘ # # Arsitektur baru (Score-Level): # LBPH → Normalize ──────────────┐ # ├─► Final Score = α·CNN_sim + β·LBPH_sim # CNN → Normalize ──────────────┤ # │ # Quality → Weight Calculator ──┘ # # Keuntungan Score-Level: # - Tidak perlu training Dense layer # - Lebih mudah dijelaskan # - Sering dipakai di sistem biometrik # - Zero retraining # # Adaptive Fusion Weight (Revisi Cross-Condition): # Masalah sebelumnya: CNN weight 70-90% terlalu dominan. # Saat foto beda kamera/kondisi, CNN sim bisa turun ke 0.37 meski LBPH bagus (0.72). # Solusi: buat weight adaptif berdasarkan CNN score aktual: # - CNN > 0.75 (kondisi bagus, kamera sama): CNN 85%, LBPH 15% # - CNN 0.55-0.75 (kondisi agak beda): CNN 65%, LBPH 35% # - CNN < 0.55 (cross-camera/kondisi berbeda): CNN 50%, LBPH 50% # # Logika: jika CNN sendiri tidak yakin, beri LBPH lebih banyak suara. FUSION_CONFIG = { 'method': 'score_level', 'use_quality_weighted': True, # Weight formula parameters (Revisi Cross-Condition: Adaptive CNN/LBPH) # CNN weight sekarang adaptif berdasarkan CNN score aktual (bukan quality saja) 'weight': { 'base_cnn': 0.65, # Base CNN weight (turun dari 0.80) 'slope': 0.10, # Slope adaptif berdasarkan quality 'min_cnn': 0.50, # Minimal CNN 50% (turun dari 70%) — cross-condition mode 'max_cnn': 0.85, # Maksimal CNN 85% (turun dari 90%) — normal mode 'min_lbph': 0.15, # Minimal LBPH 15% 'max_lbph': 0.50, # Maksimal LBPH 50% — bisa dominan saat CNN rendah }, # Threshold CNN untuk memutuskan mode fusion # Jika CNN sim < cross_condition_cnn_threshold → aktifkan LBPH rescue mode 'cross_condition': { 'enabled': True, 'cnn_low_threshold': 0.55, # CNN di bawah ini dianggap cross-condition 'cnn_high_threshold': 0.75, # CNN di atas ini dianggap kondisi bagus # CNN lebih tahan terhadap perubahan pencahayaan daripada LBPH. # Namun saat CNN rendah (cross-condition), LBPH harus bisa rescue. # Sebelumnya w_cnn_low=0.70 masih terlalu tinggi → LBPH tidak efektif. 'w_cnn_low': 0.55, # CNN rendah: lebih seimbang (55/45) 'w_cnn_high': 0.85, # kondisi bagus: CNN dominan tapi tidak ekstrem } } # ============================================================================= # 4. MULTI-FRAME EMBEDDING # ============================================================================= MULTIFRAME_CONFIG = { 'capture_frames': 15, # Untuk webcam (15 frame video) 'quality_pass_frames': 5, # Untuk webcam: ambil 5 terbaik 'photo_mode_use_all': True, # Untuk enrollment foto: gunakan semua foto } # ============================================================================= # 5. THRESHOLD CONFIG (ROC + EER, tanpa dynamic adjustment) # ============================================================================= THRESHOLD_CONFIG = { 'initial_threshold': 0.58, # Diturunkan ke 0.58 (dari 0.62) # Alasan: test cross-condition menunjukkan # fused score gelap/terang ekstrem di 0.58-0.61. # 0.58 = titik yang aman: # - Genuine cross-condition: ~0.58-0.97 ✓ # - Impostor (orang lain): ~0.35-0.55 ✗ # Kalibrasi menggunakan ROC Curve dan EER # Revisi: Hapus dynamic adjustment # Quality sudah dipakai di fusion, tidak perlu di threshold 'calibration': { 'method': 'eer', 'roc_points': 100, }, } # ============================================================================= # 6. FAISS GALLERY CONFIG # ============================================================================= FAISS_CONFIG = { 'index_type': 'IP', # Inner Product (cosine sim) 'exact_threshold': 1000, # Use exact search untuk < 1000 'top_k': 3, # Revisi: Top-3 (bukan Top-5) 'gallery_path': MODEL_DIR / "models" / "face_gallery_faiss", } # ============================================================================= # 7. INCREMENTAL TEMPLATE UPDATE # ============================================================================= # Revisi: Similarity ≥ 0.98 (lebih tinggi dari 0.90) TEMPLATE_UPDATE_CONFIG = { 'enabled': True, 'similarity_threshold': 0.98, # Revisi: 0.98 (lebih konservatif) # Exponential Moving Average 'ema_alpha': 0.2, # Weight untuk embedding baru } # ============================================================================= # 8. GALLERY QUALITY (Metadata untuk Template) # ============================================================================= # Revisi: Tambahkan Gallery Quality # Saat menyimpan embedding, simpan juga: # - Quality Score saat enrollment # - Tanggal registrasi # - Jumlah foto yang digunakan # Berguna untuk memilih template terbaik nanti GALLERY_QUALITY_CONFIG = { 'enabled': True, 'track_quality': True, 'track_date': True, 'track_count': True, } # ============================================================================= # 9. DATABASE PATHS # ============================================================================= DATABASE_CONFIG = { 'faiss_index': MODEL_DIR / "models" / "face_index.faiss", 'metadata': MODEL_DIR / "models" / "face_metadata.pkl", 'calibration': MODEL_DIR / "models" / "threshold_calibration.json", 'templates': MODEL_DIR / "models" / "face_templates.npy", } # ============================================================================= # FACE DETECTION CONFIG # ============================================================================= # Revisi Akurasi v3: Threshold dinaikkan untuk hindari false detection DETECTION_CONFIG = { 'face_confidence_threshold': 0.65, # Ditingkatkan dari 0.50 → pastikan hanya wajah jelas yang diproses 'face_nms_threshold': 0.3, 'face_max_faces': 5, 'min_face_size': 80, # Ditingkatkan dari 50 → embedding wajah kecil tidak reliable 'padding_ratio': 0.15, # Sedikit lebih besar untuk konteks } # ============================================================================= # PREPROCESSING CONFIG # ============================================================================= PREPROCESSING_CONFIG = { 'clahe_clip_limit': 2.0, 'clahe_tile_size': (8, 8), } # ============================================================================= # ENROLLMENT CONFIG # ============================================================================= ENROLLMENT_CONFIG = { 'min_frames': 1, # Minimal 1 foto (fallback) 'recommended_frames': 6, # Optimal: 6 foto per orang 'min_quality_score': 0.5, # Minimum quality per foto 'multi_template_cnn': True, # Simpan semua CNN templates individual 'use_quality_weighted_centroid': True, # Centroid tertimbang kualitas } # ============================================================================= # SINGLE PHOTO ENROLLMENT CONFIG # ============================================================================= SINGLE_PHOTO_CONFIG = { 'enabled': True, 'min_quality_score': 0.6, # Single photo perlu quality score >= 0.6 'augmentation_versions': 10, # Generate 10 augmented versions untuk stabilitas 'quality_weight_cnn': 0.75, # Higher CNN weight untuk quality yang bagus 'quality_weight_lbph': 0.25, } # ============================================================================= # AUGMENTATION CONFIG # ============================================================================= AUGMENTATION_CONFIG = { 'enabled': True, 'enrollment_copies': 10, # Lebih banyak versi untuk single photo 'max_combined': 3, # Max 3 augmentations sekaligus 'methods': { # CLAHE: clip limit lebih agresif untuk simulasi indoor 'clahe': { 'enabled': True, 'clip_limit': (2.0, 4.0), # Revisi: (1.5,3.0) → (2.0,4.0) lebih agresif 'tile_size': (8, 8), 'probability': 0.8, # Revisi: 0.7 → 0.8 (lebih sering diterapkan) }, # Gamma: lebih ekstrem untuk simulasi pencahayaan indoor gelap # Gamma < 1 → mencerahkan; Gamma > 1 → menggelapkan 'gamma': { 'enabled': True, 'gamma_range': (0.4, 1.5), # Revisi: (0.7,1.3) → (0.4,1.5) simulasi indoor 'probability': 0.7, # Revisi: 0.5 → 0.7 }, 'sharpen': { 'enabled': True, 'kernel_range': (0.5, 1.5), 'probability': 0.4, }, 'hist_eq': { 'enabled': True, 'probability': 0.3, }, # Brightness: range lebih lebar untuk simulasi gelap-terang indoor 'brightness': { 'enabled': True, 'factor_range': (0.45, 1.2), # Revisi: (0.85,1.15) → (0.45,1.2) simulasi indoor 'probability': 0.7, # Revisi: 0.5 → 0.7 }, 'rotation': { 'enabled': True, 'max_angle': 8.0, 'probability': 0.4, }, # Noise: lebih tinggi untuk simulasi grain kamera di kondisi gelap 'noise': { 'enabled': True, 'std_range': (3.0, 12.0), # Revisi: (2.0,8.0) → (3.0,12.0) simulasi noise 'probability': 0.5, # Revisi: 0.3 → 0.5 }, # Shadows: augmentasi bayangan/vignette — simulasi pencahayaan satu arah 'shadows': { 'enabled': True, 'intensity_range': (0.3, 0.7), # Intensitas shadow: 30-70% area gelap 'probability': 0.4, }, # Color Temperature: simulasi perbedaan lampu (tungsten warm vs neon cool) # Kamera Android auto white-balance sering menghasilkan color cast 'color_temp': { 'enabled': True, 'warm_shift': (10, 30), # Red boost, Blue reduce (lampu tungsten) 'cool_shift': (10, 30), # Blue boost, Red reduce (lampu neon/TL) 'probability': 0.5, }, # JPEG Compression: simulasi artefak kompresi kamera HP & WhatsApp # Foto dari Android sering di-compress ulang sebelum dikirim ke server 'jpeg_compression': { 'enabled': True, 'quality_range': (30, 70), # Quality rendah = artefak lebih banyak 'probability': 0.4, }, # Gaussian Blur: simulasi sedikit out-of-focus dari kamera HP 'gaussian_blur': { 'enabled': True, 'kernel_range': (3, 5), # Kernel kecil, blur ringan 'probability': 0.3, }, # Exposure Simulation: simulasi auto-exposure kamera Android # Kombinasi brightness + gamma + saturation shift yang realistis 'exposure_simulation': { 'enabled': True, 'underexpose_range': (0.4, 0.7), # Gelap (indoor tanpa flash) 'overexpose_range': (1.3, 1.8), # Terang (outdoor / flash) 'probability': 0.5, }, } } # ============================================================================= # LEGACY SUPPORT # ============================================================================= FACE_GALLERY_PATH = MODEL_DIR / "models" / "face_gallery.pkl" RECOGNITION_THRESHOLD = THRESHOLD_CONFIG['initial_threshold'] # ============================================================================= # HELPER FUNCTIONS # ============================================================================= def get_lbph_dim(): """Calculate LBPH feature dimension.""" return LBPH_DIM def get_reference_points(): """Get normalized reference points for face alignment.""" return ALIGNMENT_CONFIG['reference_points'].copy() # ============================================================================= # ANTI-SPOOFING CONFIG (Opsional — tidak mempengaruhi sistem existing) # ============================================================================= # Konfigurasi untuk AntiSpoofManager dan AntiSpoofGuardedSystem. # # PENTING: enabled=False by default. # Sistem existing (EnhancedFaceRecognitionSystem) TIDAK membaca config ini. # Hanya AntiSpoofGuardedSystem yang menggunakannya secara eksplisit. # # Cara mengaktifkan: # from anti_spoof_manager import AntiSpoofManager, AntiSpoofGuardedSystem # mgr = AntiSpoofManager.from_config() # system = AntiSpoofGuardedSystem(anti_spoof_manager=mgr) # ANTI_SPOOF_CONFIG = { # Aktifkan/nonaktifkan anti-spoofing. Default: False (tidak aktif). 'enabled': False, # Folder tempat menyimpan file .pth model MiniFASNet. # Buat folder ini dan letakkan file .pth di dalamnya secara manual. 'model_dir': str(_CURRENT_DIR / "resources" / "anti_spoof_models"), # Path ke hasil clone SilentFaceAntiSpoofing repository. # Clone di root proyek: git clone https://github.com/minivision-ai/Silent-Face-Anti-Spoofing SilentFaceAntiSpoofing 'spoof_lib_path': str(_CURRENT_DIR.parent / "SilentFaceAntiSpoofing"), # Model default yang digunakan (ringan, ~0.4M params, ~1.6MB). # Tersedia: 2.7_80x80_MiniFASNetV2.pth, 4_0_0_80x80_MiniFASNetV1SE.pth 'default_model': "2.7_80x80_MiniFASNetV2.pth", # Threshold probability untuk klasifikasi "REAL". # Turunkan (0.3-0.4) untuk lebih sensitif, naikkan (0.6-0.7) untuk lebih strict. 'real_threshold': 0.5, # Device inferensi: None = auto-detect (CUDA jika ada, fallback CPU). 'device': None, # Scale factor untuk crop wajah sebelum anti-spoofing. # 2.7 sesuai dengan paper MiniFASNet (context wajah lebih lebar diperlukan). 'scale_factor': 2.7, # Jika True, jalankan semua model .pth dan lakukan majority voting. # Jika False, hanya gunakan default_model. 'use_multi_model_voting': False, } def get_quality_weights(quality_score: float) -> dict: """ Calculate fusion weights based on quality score. Formula (Revisi - CNN dan LBPH Seimbang 50:50): wCNN = clip(0.50 + 0.10(Q - 0.5), 0.40, 0.60) wLBPH = 1 - wCNN Range: - CNN: 40% - 60% (seimbang) - LBPH: 40% - 60% (seimbang) Args: quality_score: Q ∈ [0, 1] Returns: Dictionary dengan w_lbph dan w_mobilefacenet """ cfg = FUSION_CONFIG['weight'] # wCNN = clip(base + slope * (Q - 0.5), min_cnn, max_cnn) w_cnn = cfg['base_cnn'] + cfg['slope'] * (quality_score - 0.5) w_cnn = max(cfg['min_cnn'], min(cfg['max_cnn'], w_cnn)) # wLBPH = 1 - wCNN w_lbph = 1.0 - w_cnn return { 'lbph': w_lbph, 'mobilefacenet': w_cnn }