skiniq - ai skincare analysis & product recommendation

computer vision pipeline that analyzes facial skin images, classifies skin type, scores key skin concerns, and recommends skincare products.

what it does

• classifies skin as oily, dry, or normal using computer vision
• extracts 6 skin metrics per image using OpenCV
• trains a Random Forest classifier on extracted features
• recommends real skincare products based on detected skin type
• generates a full skin analysis report per image

metrics analyzed

• texture score - laplacian variance of skin surface
• smoothness score - inverse texture measurement
• pigmentation score - color channel variance across RGB
• redness score - red channel ratio analysis
• pore visibility - high frequency component detection
• overall health score - weighted combination of all metrics

model performance

trained Random Forest classifier on 125 images across 3 skin types. confusion matrix included in repo showing classification breakdown.

pipeline

image → feature extraction (OpenCV) → skin type classification (Random Forest) → concern scoring → product recommendation (Sephora dataset)

results

skin type	avg smoothness	avg pigmentation	avg health
dry	13.81	78.85	18.85
normal	17.57	84.11	19.00
oily	19.20	78.67	21.05

tech stack

python · opencv · scikit-learn · pandas · matplotlib · seaborn · google colab

datasets

• oily/dry/normal skin types image dataset - kaggle
• sephora cosmetics products dataset - kaggle

key findings

• oily skin scored highest on health metrics (21.05) due to light reflectance properties captured by laplacian variance
• dry skin showed lowest smoothness (13.81) - consistent with known texture characteristics of dry skin types
• classifier achieved 32% accuracy on compressed low-res images - production accuracy would improve significantly with high-res controlled lighting images (standard in commercial skin analysis tools)
• dry vs normal misclassification mirrors challenges seen in commercial skin analysis tools - a known problem in the field
• product recommendations correctly map to skin type needs: oily → lightweight gel formulas dry → rich hydrating treatments
  normal → balanced maintenance products

limitations & next steps

• model accuracy limited by image quality - next step is fine-tuning on high resolution clinical images

• health scoring formula can be improved with dermatologist validated weights

• recommendation engine can be expanded with ingredient analysis to flag irritants for sensitive skin

  Language: Python