This presentation from "Tag der digitalen Freiheit" (Digital Freedom Day) by Christian Seiler introduces hyperspectral cameras, which perceive hundreds of colors beyond the human eye's three-color perception. The talk explains the fundamental principles of light and how these advanced cameras function, breaking down light into individual spectral components for every pixel in an image. Various types of hyperspectral cameras are discussed, with a focus on the most commonly used "Pushbroom" method and its underlying spectrograph optics. Practical applications are highlighted, including highly precise color measurement (allowing simulation under various lighting conditions from a single capture) and the non-destructive determination of active ingredient content in pharmaceuticals using Near-Infrared (NIR) spectroscopy and machine learning. The speaker also touches upon the complexities of lighting and data analysis in hyperspectral imaging.
Jenseits des Menschlichen Auges – Wie funktioniert eine Hyperspektralkamera? (Beyond the Human Eye – How does a Hyperspectral Camera work?)
0:21 Introduction: The talk introduces hyperspectral cameras, which operate beyond the visible spectrum of the human eye, and acknowledges the volunteer efforts behind the "Tag der digitalen Freiheit" conference.
1:06 Audience Engagement: A quick poll reveals that most of the audience is unfamiliar with hyperspectral cameras, confirming the talk's focus on basic principles.
1:23 Understanding Light & Spectra: The speaker reviews the electromagnetic spectrum, emphasizing that visible light is a small part. Human eyes perceive three colors, but hyperspectral cameras decompose light for each pixel into its individual spectral components, providing a complete wavelength distribution.
2:53 Complexity & Trade-offs: Building and deploying hyperspectral cameras is not simple, involving trade-offs in design.
Whiskbroom (3:29): Uses a single spectrometer to scan points individually, then compiles them into an image.
Pushbroom (4:14): Captures an entire line (row) of pixels simultaneously, each with a full spectrum. It requires movement of either the camera or the object (e.g., on a conveyor belt) to build a complete image. This is highlighted as the most commonly used type.
Mosaic Imager (4:55): Employs different filters on individual sensor pixels in a repeating pattern (e.g., 4x4), similar to color cameras but with more spectral bands (e.g., 16 vs. 3).
Filter Wheel (5:52): A standard camera with a rotating wheel of various filters, capturing multiple images sequentially to reconstruct spectral information.
6:38 How a Pushbroom Camera Works - The Spectrograph:
A spectrograph uses an entrance slit, lenses, and a grating to diffract light into its constituent wavelengths, mapping different colors to different points on a detector.
For a Pushbroom camera, this concept is extended to a slit that captures a line of light. One dimension of the 2D detector captures spatial information, while the other captures spectral information.
By moving the camera or object, a 3D "Data Cube" is built, containing spectral data for every pixel, often resulting in very large file sizes (gigabytes per acquisition).
9:52 Applications of Hyperspectral Imaging: The speaker presents a non-exhaustive list of applications.
10:02 Exact Color Measurement: Hyperspectral cameras can precisely measure how much light is reflected at each wavelength. Using standardized Tristimulus curves, this allows for mathematical simulation of color perception under any arbitrary lighting condition from just a single measurement. An example using a color preference card is shown.
14:10 Active Ingredient Analysis in Pharmaceuticals: By operating in the Near-Infrared (NIR) range, hyperspectral cameras can non-destructively determine the active ingredient content in medications. An example with paracetamol suppositories shows that while visual differences are subtle, statistical machine learning methods can accurately differentiate concentrations. This method is effective for significant concentrations, but not for trace amounts (e.g., parts per million).
16:26 Discussion on Data & Limitations: The audience inquires about the amount of data needed for training machine learning models (a 10x10 pixel area can yield 100 spectra). The speaker emphasizes the need for diverse training data to account for variations beyond just the active ingredient and the difficulty of precisely controlling lighting conditions in 3D environments.
18:13 Spectral Shape vs. Amplitude: It's clarified that for active ingredient analysis, the shape of the spectra is more critical than the amplitude, as amplitude can be influenced by the object's 3D structure and uneven illumination, often being normalized out in the model. The method primarily relies on surface reflection or shallow light penetration.
19:21 Conclusion: The speaker concludes the talk, offering to answer further questions at the conference.