Skin Type Classification with CNN

This project uses deep learning to classify human facial skin types into oily, dry, or normal, helping skincare beginners select suitable products and avoid harmful misusage.

Objective

To build a Computer Vision model that detects skin type from images, providing users with personalized skincare advice based on their actual skin condition.

Why ?

Many people misjudge their skin type, often choosing inappropriate skincare products, which can worsen issues like irritation or acne.
Source references:

• Healthshots: Dehydrated vs Oily Skin
• Canoe: Common Skincare Issues

Target Users

People who are new to skincare and unsure about their skin type.

Approach

1. Data Collection: Kaggle dataset (Oily-Dry-Normal Skin Types) split into train/validation/test.

2. EDA: Identified data quality issues like wrong labels, noise (ads, partial faces), lighting inconsistencies.

3. Preprocessing:

   • Resizing, rescaling.
   • Data augmentation: rotation, zoom, brightness control.
   • Class rebalancing and visual inspection.

4. Model 1:

   • CNN with Conv2D → MaxPooling → Flatten → Dense.
   • Issues: Overfitting, low validation accuracy (~35–40%).

5. Model 2 Improvements:

   • HeNormal kernel initializer.
   • BatchNormalization and Dropout (0.5, 0.3).
   • Learning rate: 1e-4.
   • Callbacks: EarlyStopping, ReduceLROnPlateau, ModelCheckpoint.
   • Results: Smoother loss curves, validation accuracy ↑, more balanced class performance.

6. Deployment:

   • Built Streamlit web app.
   • Users can upload an image to receive real-time skin type predictions.

Try it here: Skin Type Classification App

Results

• Validation Accuracy: ~41%
• Improved classification for “dry” and “oily” classes.
• Significant reduction in model overfitting.
• Final model saved as .keras.

⚠️ Note: The current validation accuracy (~41%) indicates this is a proof of concept. Further improvements, especially in data quality and model architecture, are ongoing.

Tech Stack

• Python
• TensorFlow / Keras
• Streamlit
• Pandas / NumPy / Matplotlib / Seaborn
• PIL / OpenCV
• scikit-learn
• Kaggle Dataset

Model Inference

Live prediction available in inference.ipynb.
Model correctly predicts skin types but still needs data cleaning and possibly transfer learning for stronger generalization.

To-Do (Next Steps)

• Clean the dataset (remove mislabeled/irrelevant images).
• Try transfer learning (e.g., MobileNetV2, EfficientNet).
• Improve UI/UX in Streamlit.
• Possibly include lighting correction or face detection.

Author

Fadhola Asandi
Skin Type Classification Project — June 2025