- Home
- Documents
*Compressed Sensing and Electron Microscopy Compressed Sensing and Electron Microscopy Peter Binev,...*

prev

next

out of 53

View

1Download

0

Embed Size (px)

Compressed Sensing and Electron Microscopy

Peter Binev, Wolfgang Dahmen, Ronald DeVore,

Philipp Lamby, Daniel Savu, and Robert Sharpley∗

Bericht Nr. 316 Dezember 2010

Key words: Compressed sensing, electron microscopy, sparsity, optimal encoding and decoding.

AMS subject classifications : 94A12, 65C99, 68P30, 41A25, 15A52

Institut für Geometrie und Praktische Mathematik

RWTH Aachen

Templergraben 55, D–52056 Aachen (Germany)

∗ This research was supported by the Office of Naval Research Contracts ONR–N00014–08–1–1113 and ONR–N00014–05–1–0715; the ARO/DoD Contracts W911NF–05–1–0227 and W911NF–07–1–0185; the NSF Grant DMS–0810869; the Special Priority Program SPP 1324, funded by DFG.

Compressed Sensing and Electron Microscopy

Peter Binev, Wolfgang Dahmen, Ronald DeVore, Philipp Lamby, Daniel Savu, and Robert Sharpley ∗

December 11, 2010

Abstract

Compressed Sensing (CS) is a relatively new approach to signal acquisition which has as its goal to minimize the number of measurements needed of the signal in order to guarantee that it is captured to a prescribed accuracy. It is natural to inquire whether this new subject has a role to play in Electron Microscopy (EM). In this paper, we shall describe the foundations of Compressed Sensing and then examine which parts of this new theory may be useful in EM.

AMS Subject Classification: 94A12, 65C99, 68P30, 41A25, 15A52

Key Words: compressed sensing, electron microscopy, sparsity, optimal encoding and decoding.

1 Introduction

Modern electron microscopic imaging has reached resolutions significantly better than 100pm which allows for unprecedented measurements of the composition and structure of materials [23, 32]. It is fair to say that imaging matter using electron microscopes, in particular STEM (scanning transmission electron microscopes, see [22], [2]), will become increasingly important in the near future, especially in biology.

However, one faces several severe obstacles to fully exploiting the information provided by aberration-corrected instruments. On the one hand, one needs to constantly remediate and reduce environmental perturbations such as air flow, acoustic noise, floor vibrations, AC and DC magnetic fields, and temperature fluctuations. On the other hand, high resolution and a good signal to noise ratio requires a high density of electrons per square nanometer. Unfortunately, soft materials are very susceptible to beam damage, and can only be visualized with low dose beams, resulting in poor resolution and a prohibitively low signal to noise ratio.

Thus, a critical issue in electron microscopy is the amount of dose needed to produce an image. Higher dose scans can damage the specimen while lower dose scans result in

∗This research was supported by the Office of Naval Research Contracts ONR-N00014-08-1-1113 and ONR-N00014-05-1-0715; the ARO/DoD Contracts W911NF-05-1-0227 and W911NF-07-1-0185; the NSF Grant DMS-0810869; the Special Priority Program SPP 1324, funded by DFG

1

high noise content in the signal. A central question is therefore how low can one keep the dose while still being able to faithfully extract the information offered by the physically possible resolution level. This calls for the development of specially tailored imaging techniques for electron microscopy that are able to go beyond the confines of currently used off-the-shelf tools.

Compressed Sensing (CS) is an emerging new discipline which offers a fresh view of signal/image acquisition and reconstruction. The goal of compressed sensing is to acquire a signal with the fewest number of measurements. This is accomplished through innovative methods for sampling (encoding) and reconstruction (decoding). The purpose of this paper is to describe the main elements of compressed sensing with an eye toward their possible use in Electron Microscopy (EM). In fact, correlating “low dose” with “fewest possible measurements” triggers our interest in exploring the potentially beneficial use of CS-concepts in EM.

In the following section, we shall give the rudiments of Compressed Sensing. We tailor our presentation to the acquisition and reconstruction of images since this matches the goals of EM. The subsequent sections of this paper will discuss possible uses of CS in Electron Microscopy. More specifically, we shall address two scenarios. The first applies to high resolution EM acquisition for materials with crystalline-like lattice structure, and the second corresponds to a much lower resolution level, which is a typical setting for electron tomography.

2 The foundations of compressed sensing

The ideas of compressed sensing apply to both image and signal acquisition and their reconstruction. Since our main interest is to discuss whether these ideas have a role to play in Electron Microscopy, we shall restrict our discussion to image acquisition.

Typical digital cameras acquire an image by measuring the number of photons that impinge on a collection device at an array of physical locations (pixels). The resulting array of pixel values is then compressed by using a change of basis from pixel representa- tion to another representation such as discrete wavelets or discrete cosines. In this new representation, most basis coefficients are small and are quantized to zero. The positions and quantized values of the remaining coefficients can be described by a relatively small bitstream.

Since the compressed bitstream uses far fewer bits than the original pixel array, it is natural to ask whether one could have - in the very beginning - captured the image with fewer measurements; for example a number of measurements which is comparable to the number of bits retained. Compressed Sensing answers this question in the affirmative and describes what these measurements should look like. It also develops a quantitative theory that explains the efficiency (distortion rate) for these new methods of sampling.

The main ingredients of this new theory for sensing are: (i) a new way of modeling real world images by using the concept of sparsity, (ii) new ideas on how to sample images, (iii) innovative methods for reconstructing the image from the samples. Each of these components can shed some light on Electron Microscopy and indeed may improve the methodology of EM acquisition and processing. To understand these possibilities we first

2

describe the primary components of CS.

2.1 Models classes for images

A digitized image is an array of N pixel values which can be represented by a matrix with real entries. We can also think of each digitized image as a vector f ∈ IRN obtained by scanning the pixel values in a specified order; usually this is the first row from left to right and then the second row left to right and so on. We shall treat the components of f as real numbers, although sensors would quantize these real numbers to a certain number of bits (typically eight or sixteen). One should view N as very large. As the resolution of sensors improves, N will grow.

If all possible vectors f ∈ IRN could appear as the pixel array of an image, there would be no hope for compression or fast acquisition. However, it is generally agreed that the images that are of interest represent a small number of the mathematically possible f . How can we justify this claim when we do not have a precise definition of real world images? We present the two most common arguments.

Firstly, one can carry out the following experiment. Randomly assign pixel values and display the resulting image. Each such image is a mathematically allowable image occurring with equal probability. One will see that all of the resulting images will have no apparent structure and do not match our understanding of real world images. Thus, real world images are such a small percentage of the mathematically possible images that we never even see one by this experiment.

A second more mathematical argument is to recognize that the pixel values that occur in a real world image have some regularity. This is not easy to see with the pixel representation of the image so we shall make a basis transformation to draw this out. The pixel representation can be thought of as representing the vector f in terms of the canonical basis functions ei ∈ IRN , i = 1, . . . , N , where the vector ei is one in the i-th position but zero in all other entries. So f =

∑N i=1 p(i)ei with p(i) the corresponding

pixel value. There are of course many other natural bases {b1, b2, . . . , bN} (with bj ∈ IRN) that could also be used to represent f . Two that are commonly used for images are the discrete Fourier and a discrete wavelet bases. We can write our image vector f in terms of these basis elements, f =

∑N i=1 x(i)bi. Notice that the coefficient vector x = Bf for a

suitable change of basis N ×N matrix B. The vector x is again in IRN . If one carries out this change of basis for real world images to either of the above mentioned bases, then one observes that most of the coefficients x(i) are zero or very small.

Figures 2.1-2.2 are an illustration of this fact. The 512×512 raw image in Figure 2.1 a) is of an M1 catalyst, a phase of mixed-metal oxide in the system Mo-V-Nb-Te-O from EM. Although this image looks to have very regular structure, a magnification of the image (Figure 2.1 b)) demonstrates that there is little regularity at the pixel level.

If we look at the histogram of pixel values there is no particular structure (Fig- ure 2.2 a)). However, if we write this image in a wavelet representation (Haar system, for example), then