AI summary by TwinMind
Lecture 8: Survey of Hyperspectral Imaging Techniques
MIT OpenCourseWare
59m 20s
Course Introduction and Creative Commons License
- MIT OpenCourseWare content under Creative Commons license
- Support enables free high quality educational resources
- Visit ocw.mit.edu for additional materials from hundreds of MIT courses
Speaker Introduction and Background
- Expert in wavelengths/spectral imaging from collaborative research background
- Built various cameras, worked on compressive sensing
- Currently works for MITRE organization
- Presenting beautiful study comparing different multispectral image capture methods
- Background spans computational photography/imaging community
- Came from physics/optics side rather than computer vision/graphics
Tomography Background Concepts
- Foundation for techniques discussed in main presentation
- Basic principle: integrate measurements along paths through objects
- Used in CAT scans (computed tomography), MRI scans
- Line integrals through objects measuring properties like density
- Multiple measurements from different angles by rotating source/detector
- Two main geometries:
- Parallel beam tomography
- Fan beam tomography (easier for point X-ray sources)
Mathematical Framework
- Parallel beam projections mathematically called Radon transform
- Each vertical slice = one projection from single angle
- Vertical slice = line through 2D Fourier transform of image
- Fourier transform sampled along radial lines based on scanning pattern
- Reconstruction challenge: non-rectangular grid prevents direct inverse FFT
- Filtered back projection algorithm addresses interpolation artifacts
Key Tomography Principle
- Perform integrals through object from multiple angles
- Provides sufficient information to determine full object structure
- Spherically symmetric objects with heterogeneous layers still reconstructable if concentric
Spectral Imaging Data Cube Concept
- 3D data structure: X, Y, and Lambda (wavelength)
- Resolution elements: NX, NY, NL (number of spectral channels)
- Terminology variations:
- Spectral: few channels (5-10)
- Multi-spectral: dozens of channels
- Hyperspectral: ~1000 channels
Comparison Framework Parameters
- Data volume: typically NX × NY × NL
- Physical device volume
- Acquisition time impact
- Computational reconstruction requirements
- Photon efficiency (major concern - light wastage)
- Compression capabilities
Baseline Method 1: Scanning Filter Camera
- Standard lens imaging with tunable wavelength filter
- Filter types:
- Color wheel with different spectral filters
- Notch filters (5-20nm bandwidth)
- Electrically tunable cavity-based filters
- Acousto-optic filters (RF-driven, ~100MHz)
- Acousto-optic filter mechanism:
- Piezoelectric transducer generates traveling pressure waves
- Light scatters off sound waves like diffraction grating
- Frequency change adjusts grating period
- Wavelength selection via aperture positioning
- Photon efficiency: 1/L (major limitation)
- Acquisition requires filter scanning over time
Baseline Method 2: Push Broom Spectrometer
- Similar to handheld spectrometer demonstration
- Light imaged onto entrance slit
- Collimated light hits diffraction grating
- Wavelength-dependent angular dispersion
- Refocused onto sensor creating vertical wavelength stripes
- Single frame captures X-Lambda slice of data cube
- Y dimension obtained through scanning (aircraft motion)
- Photon efficiency: 1/NY (light outside slit wasted)
- No computational reconstruction needed
Coded Aperture Spectrometer (Hadamard Transform)
- Architecturally identical to push broom but slit replaced with code
- Code typically half-filled, half-blocked pattern
- Collects more light than single slit (factor ~1/2 vs 1/NY)
- Hadamard codes designed for orthogonality
- Dot product operations separate wavelength contributions
- Handles spatial structure via code scanning/wrapping
- Photon efficiency: ~1/2
- Requires dot product reconstruction computations
Compressive Coded Aperture
- Same hardware as Hadamard system
- Fewer measurements than full reconstruction requires
- Underdetermined linear algebra problem: measurements < parameters
- Compressive sensing principles:
- Assume sparsity in some domain (spatial, wavelet, etc.)
- Design measurement matrix appropriately
- Use specialized reconstruction algorithms
- Poor computational scaling with data size
- Trade-off: fewer measurements for increased computation
Single vs Dual Disperser Coded Apertures
- Single disperser: modulate first, then wavelength shift
- Dual disperser: shift, modulate, then recombine wavelengths
- Trade-offs in spatial vs spectral information preservation
- Point source handling differences:
- Single: complete loss if on closed mask element
- Dual: better spatial retention, some spectral loss
Computed Tomographic Imaging Spectrometer (CTIS)
- Uses 2D diffraction grating creating multiple dispersion directions
- ~6 first-order diffraction directions plus higher orders
- Multiple gratings at 120° rotations combined
- Wavelength shifts create spatial displacement
- Mathematical analogy to medical tomography:
- X, Y, Lambda data cube instead of X, Y, Z
- Line integrals through spectral-spatial volume
- Multiple projection angles for reconstruction
- Major limitation: significant pixel waste due to spacing requirements
- Resolution loss factor of 25+ due to non-overlapping projections
Assignment 4: Multispectral Imaging
- 32-channel dataset (10 micron spacing each)
- Subjects: flowers, people, beer, etc.
- Task: mix/match spectral channels to create interesting RGB images
- Alternative: propose custom 6-hour project
- Focus: understanding spectral appearance differences
Upcoming Schedule
- Next week: scientific imaging, microscopy, tomography
- Guest speaker presentation
- Exam overview for April 13th
- Open book/laptop/everything exam format
- Primary focus should be final project preparation
