Related Experiment Video
Updated: Aug 10, 2026
07:12
Protocol for Data Collection and Analysis Applied to Automated Facial Expression Analysis Technology and Temporal Analysis for Sensory Evaluation
Published on: August 26, 2016
Do not smile when acquiring consumer selfies for multi-attribute skin profiling and baseline deep learning evaluation
Dennis Hartmann1, Dominik Müller2, Florian Auer2
1IT-Infrastructure for Translational Medical Research, University of Augsburg, Augsburg, Germany. Dennis.Hartmann@informatik.uni-augsburg.de.
Scientific Reports
|August 8, 2026
Summary
Researchers developed a new dataset and baseline models for classifying cosmetic skin features from selfies. This advances personalized skincare by enabling automated skin analysis for product recommendations.
Area of Science:
- Dermatology and Computer Vision
- Artificial Intelligence in Cosmetics
Background:
- Consumer demand for personalized skincare is high, but effective automated skin profiling for product recommendations is lacking.
- Current methods for skin analysis often require professional guidance, creating a barrier for consumers.
- Image classification offers a potential solution for objective, large-scale skin feature assessment.
Purpose of the Study:
- To create a benchmark dataset for cosmetic skin feature classification from consumer selfies.
- To establish a baseline performance evaluation for deep learning models on these skin features.
- To identify the capabilities and limitations of current deep learning architectures for automated skin analysis.
Main Methods:
- A dataset of 3203 standardized facial consumer selfies was curated and annotated for eight cosmetic skin features.
- The AUCMEDI framework was used for transparent baseline evaluation of deep learning models.
- Various deep learning architectures were trained to classify features like sagging skin, wrinkles, acne, and pigment spots.
- Mean Absolute Error (MAE) was employed to evaluate model performance, considering ordinal grading.
Main Results:
- Baseline models demonstrated promising performance for structural features such as sagging skin and wrinkles.
- Satisfactory results were achieved for under-eye circles, redness, and shine.
- Significant limitations were observed in classifying localized or imbalanced features like acne, pore size, and pigment spots.
Conclusions:
- The developed dataset and baseline provide a foundation for advancing automated cosmetic skin analysis.
- Deep learning models show potential for classifying certain skin features but require architectural improvements for others.
- Further research is needed to overcome limitations in classifying complex and subtle skin characteristics for personalized skincare.
