Super-Resolution. Dr. Yossi Rubner. Many slides from Miki Elad - Technion

Similar documents
Super-Resolution. Shai Avidan Tel-Aviv University

Simultaneous Multi-frame MAP Super-Resolution Video Enhancement using Spatio-temporal Priors

What is Image Deblurring?

ITERATED SRINKAGE ALGORITHM FOR BASIS PURSUIT MINIMIZATION

Motion Estimation (I)

Templates, Image Pyramids, and Filter Banks

Atmospheric Turbulence Effects Removal on Infrared Sequences Degraded by Local Isoplanatism

Motion Estimation (I) Ce Liu Microsoft Research New England

Sparse & Redundant Representations by Iterated-Shrinkage Algorithms

Satellite image deconvolution using complex wavelet packets

Recent Advances in SPSA at the Extremes: Adaptive Methods for Smooth Problems and Discrete Methods for Non-Smooth Problems

EE 367 / CS 448I Computational Imaging and Display Notes: Image Deconvolution (lecture 6)

Single Exposure Enhancement and Reconstruction. Some slides are from: J. Kosecka, Y. Chuang, A. Efros, C. B. Owen, W. Freeman

6.869 Advances in Computer Vision. Bill Freeman, Antonio Torralba and Phillip Isola MIT Oct. 3, 2018

Bayesian Paradigm. Maximum A Posteriori Estimation

Lecture 3: Linear Filters

Modeling Blurred Video with Layers Supplemental material

Lecture 3: Linear Filters

Computer Vision Lecture 3

Inverse problem and optimization

Image Processing in Astrophysics

Image Alignment and Mosaicing Feature Tracking and the Kalman Filter

Covariance-Based PCA for Multi-Size Data

EE Camera & Image Formation

Accelerated Dual Gradient-Based Methods for Total Variation Image Denoising/Deblurring Problems (and other Inverse Problems)

TRACKING SOLUTIONS OF TIME VARYING LINEAR INVERSE PROBLEMS

Atmospheric turbulence restoration by diffeomorphic image registration and blind deconvolution

Computer Vision & Digital Image Processing

Noise Removal? The Evolution Of Pr(x) Denoising By Energy Minimization. ( x) An Introduction to Sparse Representation and the K-SVD Algorithm

SPARSE SIGNAL RESTORATION. 1. Introduction

Camera calibration. Outline. Pinhole camera. Camera projection models. Nonlinear least square methods A camera calibration tool

LPA-ICI Applications in Image Processing

A Factorization Method for 3D Multi-body Motion Estimation and Segmentation

Motion estimation. Digital Visual Effects Yung-Yu Chuang. with slides by Michael Black and P. Anandan

Uncertainty Models in Quasiconvex Optimization for Geometric Reconstruction

Regularization methods for large-scale, ill-posed, linear, discrete, inverse problems

ENERGY METHODS IN IMAGE PROCESSING WITH EDGE ENHANCEMENT

Dictionary Learning for photo-z estimation

Science Insights: An International Journal

Wavelet Footprints: Theory, Algorithms, and Applications

CITS 4402 Computer Vision

Iterative Image Registration: Lucas & Kanade Revisited. Kentaro Toyama Vision Technology Group Microsoft Research

Preconditioning. Noisy, Ill-Conditioned Linear Systems

2 Regularized Image Reconstruction for Compressive Imaging and Beyond

Gradient-domain image processing

Computational Photography

Taking derivative by convolution

Overview. Optimization-Based Data Analysis. Carlos Fernandez-Granda

LINEARIZED BREGMAN ITERATIONS FOR FRAME-BASED IMAGE DEBLURRING

Video and Motion Analysis Computer Vision Carnegie Mellon University (Kris Kitani)

Kernel Correlation for Robust Distance Minimization

Sparse Sampling: Theory and Applications

TRACKING and DETECTION in COMPUTER VISION Filtering and edge detection

A Method for Blur and Similarity Transform Invariant Object Recognition

Nonlinear diffusion filtering on extended neighborhood

Math 56 Homework 5 Michael Downs

Galaxies in Pennsylvania. Bernstein, Jarvis, Nakajima, & Rusin: Implementation of the BJ02 methods

A RAIN PIXEL RESTORATION ALGORITHM FOR VIDEOS WITH DYNAMIC SCENES

PHASE UNWRAPPING. Sept. 3, 2007 Lecture D1Lb4 Interferometry: Phase unwrapping Rocca

From Fourier Series to Analysis of Non-stationary Signals - II

Lecture 04 Image Filtering

Feature extraction: Corners and blobs

Image Alignment and Mosaicing

The shapes of faint galaxies: A window unto mass in the universe

Notes on Regularization and Robust Estimation Psych 267/CS 348D/EE 365 Prof. David J. Heeger September 15, 1998

ERROR COVARIANCE ESTIMATION IN OBJECT TRACKING SCENARIOS USING KALMAN FILTER

Preconditioning. Noisy, Ill-Conditioned Linear Systems

Learning-Based Image Super-Resolution

Supplementary Information for cryosparc: Algorithms for rapid unsupervised cryo-em structure determination

Part III Super-Resolution with Sparsity

Edges and Scale. Image Features. Detecting edges. Origin of Edges. Solution: smooth first. Effects of noise

Affine Structure From Motion

10. Multi-objective least squares

Two-View Segmentation of Dynamic Scenes from the Multibody Fundamental Matrix

Introduction to Compressed Sensing

Image Processing in Astronomy: Current Practice & Challenges Going Forward

A new class of morphological pyramids for multiresolution image analysis

You should be able to demonstrate and show your understanding of:

Linear Inverse Problems

1. Abstract. 2. Introduction/Problem Statement

Super-Resolution of Point Sources via Convex Programming

Robust Multichannel Blind Deconvolution via Fast Alternating Minimization

Camera calibration Triangulation

Total Variation Blind Deconvolution: The Devil is in the Details Technical Report

6 The SVD Applied to Signal and Image Deblurring

Multiresolution schemes

Lecture 8: Interest Point Detection. Saad J Bedros

Compression at the source for digital camcorders.

IMPROVING THE DECONVOLUTION METHOD FOR ASTEROID IMAGES: OBSERVING 511 DAVIDA, 52 EUROPA, AND 12 VICTORIA

VID3: Sampling and Quantization

8 The SVD Applied to Signal and Image Deblurring

8 The SVD Applied to Signal and Image Deblurring

Multiresolution schemes

Sparse Subspace Clustering

Degeneracies, Dependencies and their Implications in Multi-body and Multi-Sequence Factorizations

Single-Image-Based Rain and Snow Removal Using Multi-guided Filter

Optimization-based sparse recovery: Compressed sensing vs. super-resolution

Tutorial 9 The Discrete Fourier Transform (DFT) SIPC , Spring 2017 Technion, CS Department

Sensing systems limited by constraints: physical size, time, cost, energy

1 Sparsity and l 1 relaxation

Transcription:

Super-Resolution Dr. Yossi Rubner yossi@rubner.co.il Many slides from Mii Elad - Technion 5/5/2007

53 images, ratio :4 Example - Video

40 images ratio :4 Example Surveillance

Example Enhance Mosaics

Super-Resolution - Agenda The basic idea Image formation process Formulation and solution Maximum lielihood Robust solution MAP using bilateral filter Special cases and related problems Translational Motion Interpolation / Denoising / Deblurring SR in time

Intuition For a given band-limited image, the Nyquist sampling theorem states that if a uniform sampling is fine enough ( D), perfect reconstruction is possible. D D

Intuition Due to our limited camera resolution, we sample using an insufficient 2D grid 2D 2D

Intuition However, if we tae a second picture, shifting the camera slightly to the right we obtain: 2D 2D

Intuition Similarly, by shifting down we get a third image: 2D 2D

Intuition And finally, by shifting down and to the right we get the fourth image: 2D 2D

It is trivial to see that interlacing the four images, we get that the desired resolution is obtained, and thus perfect reconstruction is guaranteed. Intuition

What if the camera displacement is Arbitrary? What if the camera rotates? Gets closer to the object (zoom)? Rotation/Scale/Disp.

There is no sampling theorem covering this case Rotation/Scale/Disp.

Further Complications Complicated motion perspective, local motion, Blur sampling is not a point operation Different effect depending on geometrical warp Noise

Super-Resolution - Agenda The basic idea Image formation process Formulation and solution Maximum lielihood Robust solution MAP using bilateral filter Special cases and related problems Translational Motion Interpolation / Denoising / Deblurring SR in time

Image Formation Scene HR Geometric transformation F Optical Blur H Sampling D Noise LR Can we write these steps as linear operators? HR D H F LR

Geometric Transformation F Scene Geometric transformation Any appropriate motion model Every frame has different transformation Usually found by a separate registration algorithm

Geometric Transformation Can be modeled as a linear operation F F F F

Optical Blur H Geometric transformation Optical Blur Due to the lens PSF Usually H H

Optical Blur Can be modeled as a linear operation H H H H

Sampling D Optical Blur Sampling Pixel operation consists of area integration followed by decimation The integration can be part of H D is the decimation only Usually D D

H PSF * PIEL H

Fill Factor Real pixel area Geometrically, the fill-factor is less than 80% for circular microlenses since the area of an unit circle is π/4 0.7854 Square-shaped micro-lenses have poor focusing capabilities

Decimation Can be modeled as a linear operation D D D o 0 0 o 0... 0 D 0

Super-Resolution - Agenda The basic idea Image formation process Formulation and solution Maximum lielihood Robust solution MAP using bilateral filter Special cases and related problems Translational Motion Interpolation / Denoising / Deblurring SR in time

Super-Resolution - Model Geometric warp Blur Decimation High- Resolution Image F I H D V Y Low- Resolution Exposures Additive Noise F N H N D N Y N Y V N { 2 0, } D H F + V, V ~ N σ n N

Simplified Model Geometric warp Blur Decimation High- Resolution Image F I H D V Y Low- Resolution Exposures Additive Noise F N H D Y N Y V N { 2 0, } DHF + V, V ~ N σ n N

The Super-Resolution Problem Y DHF + V { 0, 2 } σ, V ~ N n Given Y The measured images (noisy, blurry, down-sampled..) H The blur can be extracted from the camera characteristics D The decimation is dictated by the required resolution ratio F The warp can be estimated using motion estimation σ n The noise can be extracted from the camera / image Recover HR image

The Model as One Equation 200,000 Y Y M Y 2 N DHF DHF2 M DHFN V V + M V 2 N r resolution factor 4 MM size of the frames 0000 N number of frames 20 F H D Y ofsize ofsize ofsize ofsize ofsize 2 2 2 2 [ r M r M ] 2 2 2 2 [ r M r M ] 2 2 2 [ M r M ] 2 2 [ r M ] 2 [ M ] [60,00060,000] [60,00060,000] [0,00060,000] [60,000] [0,000] Linear algebra notation is intended only to develop algorithm

SR - Solutions Maximum Lielihood (ML): argmin N N DHF Y 2 Often ill posed problem! Maximum Aposteriori Probability (MAP) argmin DHF Y 2 { } +λa Smoothness constraint regularization

ML Reconstruction (LS) ( ) N ML Y 2 2 DHF ε Minimize: ( ) ( ) 0 ˆ 2 2 N T T T ML Y DHF H D F ε Thus, require: N T T T N T T T Y ˆ D H F DHF D H F A B B A ˆ

LS - Iterative Solution Steepest descent ˆ N T T T n+ ˆ β n F H D n ( DHF ˆ Y ) Bac projection Simulated error All the above operations can be interpreted as operations performed on images. There is no actual need to use the Matrix-Vector notations as shown here.

LS - Iterative Solution Steepest descent ˆ N T T T n+ ˆ n β F H D n ( DHF ˆ Y ) For..N ˆ n Y geometry wrap convolve with H down sample - up sample convolve with H T inverse geometry wrap F H D T D T H T F -β ˆ n+

Example HR image LR + noise 4 Least squares Simulated example from Farisu at al. IEEE trans. On Image Processing, 04

Robust Reconstruction Cases of measurements outlier: Some of the images are irrelevant Error in motion estimation, Error in the blur function General model mismatch

Robust Reconstruction Minimize: ε 2 N ( ) DHF Y ˆ N T T T n+ ˆ β n F H D sign n ( DHF ˆ Y )

Robust Reconstruction Steepest descent For..N ˆ n ˆ N T T T n+ ˆ n β F H D sign n Y ( DHF ˆ Y ) geometry wrap convolve with H down sample - sign up sample convolve with H T inverse geometry wrap F H D T D T H T F -β ˆ n+

Example - Outliers Least squares HR image LR + noise 4 Simulated example from Farisu at al. IEEE trans. On Image Processing, 04 Robust Reconstruction

Example Registration Error L 2 norm based L norm based 20 images, ratio :4

MAP Reconstruction 2 MAP N ( ) DHF Y λa{ } ε + Regularization term: 2 Tihonov cost function A T { } Γ 2 Total variation Bilateral filter A TV { } A B P P { } l P m P α l + m S l x S m y

Robust Estimation + Regularization ( ) + + P P l P P m m y l x m l N S S Y 2 α λ ε DHF Minimize: ( ) [ ] ( ) + + + P P l P P m n m y l x n m y l x m l N n T T T n n S S S S I Y ˆ ˆ sign ˆ sign ˆ ˆ α λ β DHF H D F

Robust Estimation + Regularization ˆ β N P P l + m ( ) + [ ] ( ) l m l m DHF ˆ Y λ I S S sign ˆ S S ˆ T T T ˆ n F H D sign n n+ α l P m P x y n x y n For..N geometry wrap convolve with H down sample Y - sign up sample convolve with H T inverse geometry wrap ˆ n For l,m-p..p -β ˆ + n horizontal shift l vertical shift m - - m + l sign horizontal shift -l vertical shift -m λα From Farisu at al. IEEE trans. On Image Processing, 04

Example 8 frames Resolution factor of 4 From Farisu at al. IEEE trans. On Image Processing, 04

Images from Vigilant Ltd.

Super-Resolution - Agenda The basic idea Image formation process Formulation and solution Maximum lielihood Robust solution MAP using bilateral filter Special cases and related problems Translational Motion Interpolation / Denoising / Deblurring SR in time Limitations of Super-Resolution

Special Case Translational Motion In this case H and F commute: T T HF F H H F F T H T Y DHF + V DF H + V DF Z + V Z H SR is decomposed into 2 steps. Find blur HR image from LR images non-iterative 2. Deconvolve the result using H iterative

Intuition Y DF Z + V Z H PSF* ZPIEL*PSF* Using the samples can, at most, reconstruct Z To recover, need to deconvolve Z

Step I Find Blurred HR Minimize: ε 2 ML N ( Z) DF Z Y 2 L 2 For all frames, copy registered pixels to HR grid and average [Elad & Hel-Or, 0] L For all frames, copy registered pixels to HR grid and use median [Farisu, 04]

Solution for L 2 ( ) N ML Y Z Z 2 2 DF ε Minimize: ( ) 0 2 Z Z ε ML Thus, require: N T T N T T Y P D F DF D F R Z P ˆ R Average of HR grid Diagonal, number Of occurrences per HR grid

Step II - Deblur Minimize: ( ) H Z λa{ } 2 ε + ˆ ˆ β ˆ { ( ) + { } T H signhˆ Z A ˆ n+ n n λ n n

Example 6464 LR 256256 Before deblur 256256 After deblur From Pham at al. Proc. Of SPIE-IS&T, 05. Simulated.

Related Problems Denoising (multiple frames) Y + V, V { 0, σ 2 } ~ N n Denoising (single frame) Deblurring Y + V, Y H + V, { 0, σ 2 } V ~ N n Interpolation single-image super-resolution { 0, σ 2 } V ~ N n Y DH + V, { 0, σ 2 } V ~ N n

Super-Resolution - Agenda The basic idea Image formation process Formulation and solution Maximum lielihood Robust solution MAP using bilateral filter Special cases and related problems Translational Motion Interpolation / Denoising / Deblurring SR in time

Space-Time Super-Resolution Observing events faster than frame-rate Handles: Motion aliasing Motion blur Wor and slides by Michal Irani & Yaron Caspi (ECCV 02)

Motion Aliasing The Wagon wheel effect: Slow-motion: time time time Continuous signal Sub-sampled in time Slow motion

Motion Blur

Space-Time Super-Resolution Low-resolution images video sequences: time time time High space-time resolution sequence: time time Super-resolution in space and in time.

Space-Time Super-Resolution S h (x h,y h,t h ) y t T x x y Blur ernel: PSF l S l S n Exposure time t

Example: Motion-Aliasing Input Input 2 Input 3 Input 4 25 [frames/sec]

Example: Motion-Aliasing Input sequence in slow-motion (x3): Super-resolution resolution in time (x3): 75 [frames/sec] 75 [frames/sec]

Example: Motion-Blur (simulation) Simulated sequences of fast event Long exposure-time Low frame-rate Non-uniform distribution in time Overlay of frames One low-res sequence: Another low-res sequence: And another one...

Output sequence: Output trajectory: Deblurring: Input: (x5 frame-rate) Without estimating motion of the ball! Output: 3 out of 8 low-resolution input sequences (frame overlays; trajectories):

Example: Motion-Blur (real) Frames 4 input at sequences: collision: Video Video 2 Output frame at collision: Video 3 Video 4