Deep Learning for Age Prediction and Customer Segmentation at Self-Check-In Kiosks

Project Overview

This project uses computer vision and deep learning to build an automated age verification system for self-check-in kiosks in retail stores. The model predicts a person's age and classifies them into age-based groups (e.g., senior citizens, students) to enable regulatory compliance and personalized discount offerings.

The model was trained using the UTKFace dataset and deployed in a scenario inspired by Lidl, a European supermarket chain seeking to implement real-time age detection.

---

Business Problem

Retail chains are deploying self-check-in kiosks to enhance customer experience and reduce checkout friction. However, they must still comply with regulations regarding age-restricted products (e.g., alcohol, tobacco).
Traditional manual ID checks don't fit well with self-checkout automation. A scalable solution is needed.

Additionally, the retailer wants to offer age-based discounts and promotions (e.g., senior, student, child-specific discounts). This requires an accurate and fast way to classify customers by age group.

---

Dataset Summary

• Source: UTKFace Dataset (23,991 labeled images)
• Format: Images named in the format AGE_GENDER_ETHNICITY_TIMESTAMP.jpg
• Label Extraction: Age (0–116), Gender (0 = Male, 1 = Female)
• Image Size: Resized to 64×64 and normalized

---

Problem Framing

• Age Prediction: Regression task (continuous target)
• Age Group Classification: Multi-class classification (5 age categories)
  • 0–12 → Infants & Children
  • 13–17 → Teenagers
  • 18–34 → Students & Young Professionals
  • 35–64 → Adults
  • 65+ → Seniors

---

Workflow Summary

1. Data Acquisition & Cleaning

   • Downloaded and cleaned image files
   • Parsed filenames to extract age and gender
   • Converted images to pixel arrays using OpenCV

2. Feature Engineering

   • Created age group labels for classification
   • Normalized pixel values
   • Resized all images to 64×64×3

3. Model Building

   • Built custom Convolutional Neural Network (CNN)
   • Trained two models:
     • Regression model for age prediction
     • Classification model for age category

4. Evaluation

   • Used MAE, MSE for regression
   • Used accuracy, confusion matrix, and classification report for classification

---

Tools & Technologies

• Python
• TensorFlow / Keras
• OpenCV
• NumPy, pandas
• matplotlib, seaborn
• Jupyter Notebook

---

Key Outcomes

• Built an end-to-end deep learning pipeline for facial image-based age prediction
• Achieved accurate age classification into business-relevant categories
• Enables use in retail environments for:
  • Automated age verification (regulatory compliance)
  • Age-based discount offers (targeted marketing)

---

Author

Lohith Basavanahalli Anjinappa