A Comparison of Quantitative Image Quality Evaluation Techniques for
Transmission X-Ray Microscopy
Peter Bolgert
Marquette University
Mentor: Yijin Liu
Introduction
Beam Line 6-2c at the Stanford Synchrotron Radiation Lightsource (SSRL) is capable of
Full-Field Transmission X-ray Microscopy (TXM) at 30 nm spatial resolution. Users from
around the world have used this microscope to image a variety of biological and inorganic
samples. Depending on the specific TXM technique being used, the user will typically take
thousands of images of a single sample. These images must be extensively processed before
presentation-quality images can be obtained. While the details of the image processing depend
on the specific technique TXM used, a simplified workflow is shown in Figure 1.
With such immense amounts of data, it is necessary to automate the image processing
workflow as much as possible, especially for the alignment of images with each other. Not only
is it tedious to align images by eye, but a computer should be able to align images much more
accurately than a human. Image alignment is a recurring part of our workflow, as it appears in
steps two (Aligning and Averaging of Multiple Images), three (Mosaic Stitching), and five
(Tomographic Reconstruction).
By automating this task, users can focus more on their
experiment and less on repetitive image processing.
One popular algorithm for automatic image alignment is known as phase correlation,
which manipulates the two images in the frequency domain. The exact details of this method are
not important for this paper. The important point is that for two identical noiseless images which
differ by a pure translation, phase correlation will align the image perfectly (to the nearest pixel).
For images with high levels of noise, sometimes the phase correlation algorithm fails. For
example, sometimes the algorithm will attempt to align noise with noise, which is an
unsatisfactory result. For example, typically the user will take 20 – 30 repeated images of the
sample at the same angle. In between every exposure, the sample stage will move slightly,
causing these repeated exposures to be identical except for a small translation. Step two of our
workflow (Aligning and Averaging of Multiple Images) consists of aligning these repeated
images and then taking the average, which increases the signal-to-noise ratio. If the phase
correlation algorithm aligns noise with noise, the resulting average will be blurred.
To remedy this problem (especially in images with a lot of high frequency noise), it is
helpful to slightly blur the images before attempting to align them. This will mask the noise
while keeping the important large scale features intact, which improves the alignment. This
blurring process requires manual user input and slows down the workflow. In order to make the
process fully automatic we need an algorithm to evaluate image quality. If the averaged image is
not sharp enough, the algorithm should adjust the blurring so that the alignment is optimized.
All of this should occur without any input from the user. In this paper, we will apply 3? (not sure
how many yet) image sharpness algorithms to both real and artificial images in an attempt to
more fully automate our image processing workflow.
Materials and Methods
(since I started my project 3 weeks late, I do not have much to write in this section yet)
Figure 1: Image Processing Work Flow