Facial Attribute Classification with Custom CNN

Multi-class facial attribute classification using a custom CNN architecture with focal loss and class-weighted training to handle severe class imbalance in the UTKFace dataset.

Architecture

Custom CNN with 4 convolutional blocks:

Input(128×128×3)
→ Conv(64) → BN → ReLU → MaxPool
→ Conv(128) → BN → ReLU → MaxPool
→ Conv(256) → BN → ReLU → MaxPool
→ Conv(512) → BN → ReLU → GlobalAvgPool
→ Dense(512) → Dropout → Dense(5, Softmax)

Dataset

UTKFace — 20,000+ face images with demographic attribute labels.

• 5 classes (imbalanced distribution)
• Input resolution: 128×128
• Train/Val split: 80/20

Training Strategy

Two techniques applied to address class imbalance:

Focal Loss (Lin et al., 2017) — down-weights easy examples, focuses training on hard minority-class samples:

FL(p_t) = -α_t (1 - p_t)^γ · log(p_t)   # γ=3.0

Class-weighted loss — manual weight scaling for underrepresented classes (up to 5× for rarest class).

Hyperparameter	Value
Optimizer	AdamW
Epochs	15
Batch size	8
Input size	128×128
Augmentation	RandomFlip, Brightness, Contrast

Usage

Open in Colab. Mount Google Drive and set zip_file_path to your UTKFace dataset location. Run all cells.

References

• Lin, T. et al. (2017). Focal Loss for Dense Object Detection. ICCV 2017
• Zhang, Z. et al. (2017). Age progression/regression by conditional adversarial autoencoder. CVPR 2017 (UTKFace dataset)