The Promise of Machine Learning

Chandrajit Bajaj

1 Use visuals to make the invisible visible
2 Eye is not enough, infared data can find the fake plant
3 HyperSpectral Image (HSI) (分解影象到不同顏色空間？)
4 Diversity of Hyperspectral Imaging (Wavelength)
5 Color seperation
6 Spectrum Analysis of Multi-Spectral Data (PCA)

• Covariance matrix
• Primal Formulation
• Dual Formulation
• SVD

7 LIDAR, spatial, spectral. tensor decomposition (CP, Tucker) [http://cvcweb.ices.utexas.edu](Computational Visualization Center)
8 Top 3 PCA Approximation (green > red > blue)
9 Recover from corrupted spectral data, similar to before, but not non-linear
10 Prior Database of Spectral Images (Sparse regression from Database)
y = y* + b(large magnitude corruption) + (noise)
y* = Xw*
11 Uncorrupted Pixel Recovery & Sparse Regression
12 Stain Image Biopsy of Canderous Tissue (Manual Detection), lack of standardization for biomarker assessment
13 Modern Stainfree Biopsies (interferometer light source) (如何自動標定黑白醫學成像到有意義的彩色影象)
14 Invisible Disease to Visible Knowledge (Feature -> Digital Staining, Classification)
15 Classification fails because of noise, Signal and noise maybe correlated, Noise correlated across multiple channel
16 Maxmize the signal to Noise Ratio (使有效訊號除以噪聲最大)
17 Spectral Optimization for Minimum Noise Fraction
18 Automatic Denoised Band Selection
19 Tumor Identification & Microenvironment
20 Spectral UnMixing : H&E stained image, Sub-Pixel Classified Image
21 Remote sensing
22 Specral Mixing
23 Unmixing a Linear Mixture Model:
Robust Non-Negative Matrix Factorization
X = EA, Latent Fator Discovery
Non-Convex Optimization (Latent Factor Estimation)
Volume minimization as an approach
24 Hyperspectral Fusion
25 Tensor Fusion for Enhanced Spatial Resolution
26 Digital Stain: Transferring the Stained Image to RTIR
27 Coupled Sparse Tensor Factorization and Spatio-Spectral Super-Resolution
28 Decompose each spectral image, aligned and fused
29 Multi-Spetral and Hyperspectral image Fusion
30 CRANIO_FACIAL Templating
31 Target Shape + Shape Database -> Robust shape recovery
32 Tetralogy of Fallot, Pulmonary Valve Insufficient, Fetal Cardiac MR
33 4D: Aneurysm of Coarctation Repair

收穫:
1 分解的思想
2 同時優化訊號和噪聲的做法