Smartphone Pigment ID with PCA

ISEF Category: Chemistry

Subcategory: Other  ·  Difficulty: Intermediate  ·  Setup: School Lab  ·  Time: 1 to 2 Months

The Hook

A fake pigment can look right to your eye and still be wrong to a lab. That matters for art conservation, museum authentication, and forgery checks. You can test that problem with a phone camera, color calibration, and public spectral data.

What Is It?

This project asks a simple question with real museum stakes: can you tell an old pigment from a look-alike without damaging the sample? Pigments are the colored powders in paint. Some have clear spectral fingerprints, which means they reflect light in patterns that act like a chemical signature.

Your phone does not directly measure a full spectrum the way a lab spectrometer does. But a camera still records RGB color, and with the right calibration model, you can estimate more detailed spectral information from those RGB values. Think of it like hearing a song from only three notes, then using a trained pattern model to guess the rest. That is the core idea behind hyperspectral-from-RGB reconstruction.

Then you use PCA, which stands for principal component analysis. PCA compresses many spectral measurements into a smaller set of patterns that separate pigments that look similar in photos but differ in their underlying reflectance. In this project, you would compare historical pigments such as ochre, malachite, and lapis substitutes against mimic materials using a public pigment-spectra dataset and your own smartphone images.

Why This Is a Good Topic

This is a strong science fair topic because you can measure something real, compare classes of materials, and build a model that makes a clear prediction. You also connect chemistry to art conservation, cultural heritage, and fraud detection. A student can realistically collect image data, clean a public dataset, test a classifier, and report accuracy, confusion patterns, and limits of the method.

Research Questions

• How does smartphone camera calibration affect the accuracy of pigment classification?
• What is the effect of using PCA on reconstructed spectra versus raw RGB values for separating pigment mimics?
• Does adding different lighting conditions change how well the classifier identifies ochre, malachite, and lapis substitutes?
• To what extent can a public pigment-spectra dataset train a model that works on smartphone images of real samples?
• Which color channels or reconstructed spectral bands contribute most to distinguishing historical pigments from their mimics?
• How does sample surface texture affect the classifier's false positives for visually similar pigments?
• What is the effect of different white-balance settings on spectral reconstruction quality and class separation?

Basic Materials

• Smartphone camera with manual exposure and white balance controls.
• Color calibration target or printed color checker.
• Stable desk lamp with consistent daylight-like output.
• Neutral gray or black background.
• Samples or swatches of pigment-like materials, such as ochre, malachite-like green mineral paint, and lapis-like blue substitutes.
• Spreadsheet software for data logging.
• Computer with internet access.
• ImageJ for image measurement and channel extraction.

Advanced Materials

• Handheld or bench spectrometer for validation measurements.
• Standard reflectance targets for calibration.
• Controlled light box with high-CRI lighting.
• Raw image capture app or camera app that saves unprocessed files.
• Reference pigment samples from a teaching collection or conservation lab.
• Python environment with scientific libraries.
• PCA and classification workflow tools.
• Microscope or stereomicroscope for documenting particle texture and surface differences.

Software & Tools

• ImageJ: Measures RGB values from calibration targets and pigment regions in your photos.
• Python: Runs PCA, builds classifiers, and compares model performance across sample groups.
• Jupyter Notebook: Keeps your analysis steps organized and easy to explain.
• scikit-learn: Provides PCA, train-test splits, and classification tools for your model.
• LibreOffice Calc: Helps you log sample metadata and track calibration results.

Experiment Steps

1. Define which pigment mimics you will compare and what counts as a correct identification.
2. Choose one imaging setup, then lock down lighting, camera settings, and background so your data stay consistent.
3. Build a calibration workflow that converts phone color readings into a more stable numerical signal.
4. Match your images to the public pigment-spectra dataset and decide how you will align or normalize the data.
5. Test PCA as a feature reduction step, then compare classifier performance against raw RGB features.
6. Plan validation with held-out samples or repeated images so you can measure whether the method generalizes.

Common Pitfalls

• Changing light sources between image sessions, which makes the same pigment look like a different one in the model.
• Mixing glossy and matte samples, which changes reflections and confuses the spectral reconstruction.
• Training on too few pigment examples, which leads to a classifier that memorizes samples instead of learning pigment patterns.
• Skipping calibration targets, which makes your RGB values drift and breaks comparison across days.
• Treating reconstructed spectra as exact spectra, which can hide error from the camera and make conclusions too strong.

What Makes This Competitive

A class-level version of this project only shows that a phone can sort colors. A stronger version tests whether the method still works across lighting changes, surface textures, and mimic pigments that were chosen because they are hard to tell apart. You can raise the quality by reporting confusion matrices, cross-validation results, and error bars, then comparing raw RGB, reconstructed spectra, and PCA features. The best projects also explain when the method fails, because that tells museums and conservators where the tool can and cannot help.

Project Variations

• Compare natural earth pigments against synthetic historical substitutes to see which pairs are hardest to separate.
• Test whether adding texture features from microscope images improves classification beyond color data alone.
• Focus on one pigment family, such as blue pigments, and compare several mimic recipes or brands.

Learn More

• PubMed: Search review articles on pigment identification, reflectance spectroscopy, and conservation science.
• NIH PMC: Find free full-text papers on chemometrics, PCA, and color-based classification.
• NOAA National Centers for Environmental Information: Review documentation on spectral data handling and calibration concepts.
• NASA Open Science Data Repository: Explore open workflows for image calibration and remote sensing style spectral analysis.
• MIT OpenCourseWare, Principles of Chemical Science: Review the chemistry basics behind color, light absorption, and molecular structure.
• Journal of Raman Spectroscopy: Search for open papers on pigment identification and non-destructive analysis.