optimizing and learning for super-resolution lyndsey c. pickup, stephen j. roberts & andrew...
TRANSCRIPT
Optimizing and Learning forSuper-resolution
Lyndsey C. Pickup, Stephen J. Roberts
& Andrew Zisserman
Robotics Research Group, University of Oxford
The Super-resolution Problem
Given a number of low-resolution imagesdiffering in: geometric transformations lighting (photometric) transformations camera blur (point-spread function) image quantization and noise.
Estimate a high-resolution image:
Low-resolution image 1
Low-resolution image 2
Low-resolution image 3
Low-resolution image 4
Low-resolution image 5
Low-resolution image 6
Low-resolution image 7
Low-resolution image 8
Low-resolution image 9
Low-resolution image 10
Super-Resolution Image
Generative Model
Registrations, lighting and
blur.
High-resolution image, x.
y1 y2 y3 y4
Low-resolution images
W4W3W2W1
Generative Model
• Geometric registrations• Point-spread function• Photometric registrations
We don’t have:We have:• Set of low-resolution input images, y.
Maximum a Posteriori (MAP) Solution
Standard method:
1. Compute registrations from low-res images. 2. Solve for SR image, x, using gradient descent.
y1 y2 y3 y4
W4W3W2
W1
x
[Irani & Peleg ‘90, Capel ’01, Baker & Kanade ’02, Borman ‘04]
What’s new
1. We solve for registrations and SR image jointly.
2. We also find appropriate values for parameters in the prior term at the same time.
Hardie et al. ’97: adjust registration while optimizing super-resolution estimate.
• Exhaustive search limits them to translation only. • Simple smoothness prior softens image edges.
i.e. given the low-res images, y, we solve for the SR image x and the mappings, W simultaneously.
y1 y2 y3 y4
W4W3W2
W1
x
Overview of rest of talk
• Simultaneous Approach– Model details
– Initialisation technique
– Optimization loop
• Learning values for the prior’s parameters
• Results on real images
Maximum a Posteriori (MAP) Solution
Image x. Corrupt with additive Gaussian noise.
Warp, with parameters Φ.
Blur by point-spread function.
Decimate by zoom factor.
y1 y2 y3 y4
W4W3W2
W1
x
y
Details of Huber Prior
Huber function is quadratic in the middle, and linear in the tails.
Probability distribution is like a heavy-tailed Gaussian.
ρ (z,α) p (z|α,v)
Red: large αBlue: small α
This is applied to image gradients in the SR image estimate.
Details of Huber Prior
Ground Truth
α=0.1 v=0.4
Too little smoothingToo much smoothing
α=0.05 v=0.05 α=0.01 v=0.01 α=0.01 v=0.005
Edges are sharper
Advantages: simple, edge-preserving, leads to convex form for MAP equations.
Solutions as α and v vary:
Advantages of Simultaneous Approach
Learn from lessons of Bundle Adjustment: improve results by optimizing the scene estimate and the registration together.
Registration can be guided by the super-resolution model, not by errors measured between warped, noisy low-resolution images.
Use a non-Gaussian prior which helps to preserve edges in the super-resolution image.
Overview of Simultaneous Approach
1. Start from a feature-based RANSAC-like registration between low-res frames.
2. Select blur kernel, then use average image method to initialise registrations and SR image.
3. Iterative loop: Update Prior Values Update SR estimate Update registration estimate
Use average image as an estimate of the super-resolution image (see paper).
Minimize the error between the average image and the low-resolution image set.
Use an early-stopped iterative ML estimate of the SR image to sharpen up this initial estimate.
Initialisation
Average image
ML-sharpened estimate
1.Update prior’s parameter values (next section)
2.Update estimate of SR image
3.Update estimate of registration and lighting values, which parameterize W
Repeat till converged.
Optimization Loop
Joint MAP ResultsD
ecre
asin
g p
rio
r st
ren
gth
Registration Fixed Joint MAP
Learning Prior Parameters α, ν
Split the low-res images into two sets:
Use first set to obtain an SR image.
Find error on validation set.
Learning Prior Parameters α, ν
Split the low-res images into two sets:
Use first set to obtain an SR image.
Find error on validation set.
But what if one of the validation images is mis-registered?
Learning Prior Parameters α, ν
Instead, we select pixels from across all images, choosing differently at each iteration.
We evaluate an SR estimate using the unmarked pixels, then use the forward model to compare the estimate to the red pixels.
Learning Prior Parameters α, ν
Instead, we select pixels from across all images, choosing differently at each iteration.
We evaluate an SR estimate using the unmarked pixels, then use the forward model to compare the estimate to the red pixels.
Learning Prior Parameters α, ν
To update the prior parameters:
1. Re-select a cross-validation pixel set.
2. Run the super-resolution image MAP solver for a small number of iterations, starting from the current SR estimate.
3. Predict the low-resolution pixels of the validation set, and measure error.
4. Use gradient descent to minimise the error with respect to the prior parameters.
Results: Eye Chart
MAP version: fixing registrations then
super-resolving
Joint MAP version with adaptation of prior’s
parameter values
Results: Groundhog Day
The blur estimate can still be altered to change the SR result. More ringing and artefacts appear in the regular MAP version.
Results: Groundhog Day
Blur radius = 1 Blur radius = 1.4 Blur radius = 1.8
Regular MAP
Simultaneous
Lola Rennt
Real Data: Lola Rentt
Real Data: Lola Rentt
Real Data: Lola Rentt
Real Data: Lola Rentt
Conclusions
• Joint MAP solution: better results by optimizing SR image and registration parameters simultaneously.
• Learning prior values: preserve image edges without having to estimate image statistics in advance.
• DVDs: Automatically zoom in on regions with a registrations up to a projective transform and with an affine lighting model.
• Further work: marginalize over the registration – see NIPS 2006.