Human Age Prediction (Computer Vision)

Table of Contents

• Overview
• Business Problem
• Dataset
• Exploratory Data Analysis (EDA)
• Modeling Approach
• Results & Model Performance
• Key Insights & Business Impact
• Next Steps & Potential Improvements
• How to Use
• Connect With Me

---

Overview

This project explores computer vision techniques to predict a person’s age from images. The goal is to determine whether machine learning can help a supermarket chain ensure compliance with alcohol sales laws by verifying a customer's age at checkout.

Business Problem

Good Seed, a supermarket chain, wants to implement an automated system to prevent underage alcohol sales. They aim to:

• Use computer vision to estimate a customer’s age from a photo.
• Train a convolutional neural network (CNN) to make accurate predictions.
• Evaluate the feasibility of AI-powered age verification.

Dataset

The dataset is sourced from ChaLearn Looking at People and contains:

• Images: Stored in the final_files/ folder.
• Labels: A CSV file (labels.csv) with:
  • file_name → Image filename.
  • real_age → True age of the individual.

Given the large dataset size (~7,600 images), data generators were used for efficient loading.

---

Exploratory Data Analysis (EDA)

Key Findings

• Age distribution: The dataset covers a broad age range but has more images of younger individuals.
• Visual inspection: Sample images confirmed variability in lighting, angles, and expressions, impacting model accuracy.

Age Distribution

Sample Images Across Ages

---

Modeling Approach

The model uses a pretrained ResNet50 architecture with fine-tuning:

• Base Model: ResNet50 (pretrained on ImageNet).
• Additional Layers:
  • GlobalAveragePooling2D for feature extraction.
  • Dense layer with ReLU activation.
  • Output layer predicting a single continuous value (age).
• Loss Function: Mean Squared Error (MSE).
• Optimizer: Adam (learning rate: 0.0005).

Model Training

The model was trained on GPU using data augmentation:

• Horizontal flips and rotations to improve generalization.
• Mini-batches for memory efficiency.
• Early stopping to prevent overfitting.

---

Results & Model Performance

Epoch	Train MAE	Validation MAE
1	7.43	8.49
5	5.59	8.21
10	4.14	6.96
15	3.67	6.78
20	3.17	7.65

• The final MAE (Mean Absolute Error) = 7.1 years.
• The model performs well for middle-aged individuals but struggles with extreme age groups.
• Validation loss fluctuated, suggesting further tuning is needed.

---

Key Insights & Business Impact

• Moderate accuracy: The model can estimate age within ~7 years, making it unreliable for strict legal compliance.
• Practical applications: Useful for pre-screening but requires human verification for final decisions.
• Challenges:
  • Poor performance on very young & elderly individuals.
  • External factors (lighting, facial expressions) impact predictions.

Business Recommendations

• Use AI as an assistive tool, not a standalone verification system.
• Combine with ID scanning for more reliable validation.
• Train on a more balanced dataset with improved age representation.

---

Next Steps & Potential Improvements

• Fine-tune model hyperparameters to reduce MAE.
• Experiment with deeper architectures (EfficientNet, Vision Transformers).
• Incorporate additional facial features for better predictions.
• Train on a larger, more diverse dataset to improve generalization.

---

How to Use

Clone the repository:

  git clone https://github.com/rhi-222/human-age-prediction.git

Install dependencies:

  pip install tensorflow keras matplotlib numpy pandas seaborn

Run the Jupyter Notebook:

• Open [updated]HumanAge(ComputerVision).ipynb in Jupyter Notebook or Google Colab.
• Execute the notebook to preprocess images, train the model, and evaluate predictions.

Connect With Me

• Email: [email protected]
• LinkedIn: linkedin.com/in/rhiannonfilli