Gender Detection Model on vgg19

Gender Detection using VGG19 and Metrics

This model performs binary gender classification (Male/Female) on face images using a fine-tuned VGG19 architecture. The model is trained on the class-imbalanced FACECOM dataset using advanced techniques to ensure fairness and high generalization.

---

🧠 Model Architecture

• Base model: VGG19 (pretrained on ImageNet)
• Top layers:
  • GlobalAveragePooling2D
  • Dense(512, activation='relu')
  • Dropout(0.5)
  • Dense(1, activation='sigmoid')
• Input size: 250 × 250 × 3
• Output: Probability of class “Female” (label 1)

---

⚙️ Training Strategy

Phase	Description
1️⃣	Train only the new head layers (VGG19 frozen)
2️⃣	Fine-tune top convolutional blocks of VGG19
✅	EarlyStopping, ReduceLROnPlateau, and ModelCheckpoint for best performance

---

🧪 Techniques Used

• Loss Function: Binary Crossentropy with Label Smoothing
• Optimizer: Adam with learning rate scheduling
• Augmentation: Horizontal flip, rotation, brightness/contrast jitter
• Class Imbalance Handling:
  • Augmented female class
  • Class weights during training
• Metrics Tracked: Accuracy, Precision, Recall, F1-Score, AUC

---

📊 Performance (Validation Set)

Metric	Value
Accuracy	97%
Precision	0.9714
Recall	0.9653
F1-score	0.9684
AUC	> 0.96

---

📁 Model Files

File	Description
vgg19_train.h5	Final fine-tuned model ready for inference (recommended)
vgg19_final_epoch.h5	Final saved model & ready for inference

---

Dataset structure

Data/
└── Task_A/
    ├── train/
    │   ├── male/
    │   └── female/
    └── val/
        ├── male/
        └── female/

Embedding Model Resnet50 on Common Faces

Embedding Model of Facial Classification Triplet Network

This model performs face verification using a Triplet Network with a ResNet50 backbone, trained on the distorted face images provided in the FACECOM dataset.

🔧 Architecture Details

• Backbone: ResNet50 (pretrained on ImageNet)
• Embedding dimension: 256
• Triplet Loss: Euclidean distance with margin = 0.3
• Normalization: L2 normalized embeddings
• Head: Dense(256) + Lambda(L2 norm)
• Training strategy:
  • Phase 1: Train frozen ResNet base
  • Phase 2: Fine-tune top ResNet layers
• Input size: 224 × 224 × 3

📁 Model Files

File	Description
tripletNetwork.h5	Full Triplet Network (used for training)
embedding_sequel.h5	Embedding extractor model (used for inference)

📊 Performance (on Validation Set)

• ✅ Accuracy: ~96–97%
• ✅ Precision: 0.9729
• ✅ Recall: 0.9841
• ✅ F1 Score: 0.9785
• 📐 Threshold: 0.945

Dataset Structure

Data/
└── Task_B/
    ├── train/
    │   └── 001_identity/
    │       ├── 001_identity.jpg
    │       └── distortion/
    │           ├── 001_identity_blurred.jpg
    │           └── ...
    └── val/
        └── ...