aafreen06/FaceDetection

Face Recognition Model

A CNN-based face recognition model built from scratch using Keras/TensorFlow.

People it recognizes

• Aafreen
• Syeda
• Taha

Model Architecture

• 4 Convolutional Blocks (Conv2D → BatchNorm → ReLU → MaxPool)
• Filters: 32 → 64 → 128 → 256
• Dense(256) → Dropout(0.5) → Dense(3, Softmax)
• Input size: 128×128×3

Training Details

• Dataset: ~71 images (22–26 per person)
• Augmentation: 7 variants per training image (flip, rotation, brightness, zoom)
• Split: 70% train / 15% val / 15% test
• Optimizer: Adam (lr=0.001)
• Loss: Categorical Crossentropy
• Callbacks: EarlyStopping, ReduceLROnPlateau, ModelCheckpoint

Files

File	Description
face_model.h5	Trained Keras model
class_names.json	Label index mapping
training_curves.png	Accuracy & loss plots
confusion_matrix.png	Evaluation results

How to use

from tensorflow.keras.models import load_model
import json, numpy as np
 
model = load_model('face_model.h5')
with open('class_names.json') as f:
    class_names = json.load(f)
 
# Predict on a 128x128 face crop
img = img / 255.0
img = np.expand_dims(img, axis=0)
pred  = model.predict(img)
label = class_names[str(np.argmax(pred))]
conf  = np.max(pred)
print(f"{label} ({conf*100:.1f}%)")

Project

Applied AI Final Project — COMP 6721 Concordia University, Winter 2026