1. Image Stitching for Optical Microscopy
Timothy Blattner, Bertrand Stivalet, Walid Keyrouz
Shujia Zhou
IAB

2. Image Stitching for Optical Microscopy
Objectives
Stitching of optical microscopy images at interactive rates
General purpose library, ImageJ/Fiji plug-in, etc.
Success criterion
Transformative impact
Run sample problem in < 1 min
> 10x speed improvement
IAB

3. Credits
Joe Chalfoun, Mary Brady
NIST
IAB

4. Image Stitching Problem
Optical microscopes scan a plate and take overlapping partial images (tiles)
Need to assemble image tiles into one large image
Modern microscopy automated:
Scientists are acquiring & processing large sets of images
IAB

5. Image Stitching Problem…
Header
Image Stitching Problem…
Two phases:
Compute the X & Y translations for all tiles
Apply the translations & compose the stitched image
Main focus is on phase I
IAB
Footer

6. Image Stitching Algorithm
Loop over all images:
Read an image tile
Compute its FFT-2D
Compute correlation coefficients with west and north neighbors
Depends on FFT-2D for each tile
Major compute portions:
FFT-2D of tiles
Compute and normalize phase correlation
Inverse FFT-2D
Reduce max normalize
IAB

7. Algorithm’s Parallel Characteristics
Almost embarrassingly parallel
Large number of independent computations
For an n x m grid:
FFT for all images nxm
NCC for all image pairs 2nxm - n - m
FFT-1 for the NCCs of all image pairs 2nxm - n - m …
Caveats
Data FFT dependencies
Limited memory
IAB

8. Data Set Grid of 59x42 images (2478)
1392x bit grayscale images (2.8 MB per image)
~ 7 GB
Source:
Kiran Bhadriraju (NIST)
IAB

9. Evaluation Platform Hardware
Dual Intel® Xeon® E-5620 CPUs (quad-core, 2.4 GHz, hyper-threading)
24 GB RAM
Dual NVIDIA® TeslaTM C2070 cards
Reference Implementations
Fiji™ Stitching plugin, >3.6 hours
MATLAB® prototype, ~17.5 minutes on a similar machine
Software
Ubuntu Linux 12.04/x86_64, kernel 3.2.0
Libc , libstd
BOOST 1.48, FFTW 3.3, libTIFF4
NVIDIA CUDA & CUFFT 5.0
IAB

10. Implementations & Results FFTW Exhaustive, CUDA 5.0
Time
Speedup
CPU
Threads
GPUs
C++ Sequential
10 min 37 sec 1
Simple Multi-Threaded
1 min 48 sec
5.8x 8
Pipelined Multi-Threaded
1 min 22 sec
7.7x 19
Simple GPU
9 min 47 sec
1.08x
Pipelined-Hybrid
25 sec
25.5x 13 2
IAB

11. Java Implementation Allows easy integration info Fiji
Tool used by many biologists for image stitching
Pure Java code is extremely slow
FFT computations
Cross correlation
Use JNI with FFTW and C code
Java native interface
Allows calling functions off of the virtual machine
Requires compilation (gcc)
IAB

12. Java Implementation Runtimes 42x59 Tiles
Threads
Sequential
> 4 hours 1
Sequential with JNI
~30 minutes
Pipelined with JNI
3 min 42 sec 16
IAB

13. Closure—General 25x speedup compared to Sequential C++ code
518x speedup compared to Fiji stitching plugin
Representative data set:
42x59 grid
~25 sec
Can budget compute time to:
Generate stitched image
Carry out additional analysis
Enables computationally steerable experiments
IAB

14. Closure—Java Implementation
Single threaded-executes full grid in ~45 minutes using FFTW native interface
Multi-threaded executes in ~4 minutes
Optimized version uses native intrinsics for computing cross correlation
Provides simple integration into Fiji application
Need to provide JNI for GPU functions
JCUDA
IAB

15. Thank You Questions?
IAB