Linear Imaging Systems

Example: The Pinhole camera

Outline
General goals, definitions Linear Imaging Systems Object function Superposition Instrument response function Delta function Isoplanatic maps Point spread function Magnification Pinhole imager

Imaging Definitions

Object function - the real space description of the actual object. Resolution - the collected image is only an approximation of the actual object. The resolution describes how accurate the spatial mapping is. Distortions - describes any important nonlinearities in the image. If there are no distortions, then the resolution is the same everywhere. Fuzziness - describes how well we have described the object we wish to image. Contrast - describes how clearly we can differentiate various parts of the object in the image. Signal to Noise ratio

Imaging Definitions
There are two very basic problems in image analysis: 1. 2. Given the data, an estimation of the instrument function and a (statistical) model of the noise, recover the information (an estimation of the object function), Employing a suitable model, interpret this information.

So here our goals are: 1. Design an algorithm to compute the object function given certain data on the scattering field, an estimation of the point spread function, and an estimation of the experimental noise. 2. Find a way of recovering an appropriate set of material parameters from the object function, 3. Set up the experiment such that object function reflects the parameter of interest, 4. Design the experiment such that the imaging experiment closely approximates a linear system (or that the non-linearities are dealt with correctly) and such that the point spread function is narrow and of definite shape.

Linear Imaging Systems

A model of the imaging process is needed to extract spatial information from a measured signal. For most of this course we will be concerned with a deceptively simple linear model Image = Object Function Point Spread Function + Noise This is the basis expression for linear imaging. The Point Spread Function depends on the type and properties of the imaging system, and the Object Function depends on the physical interactions of the object and the scattering wave. The noise is an important consideration since it limits the usefulness of deconvolution procedures aimed at reversing the blurring effects of the image measurement. If the blurring of the object function that is introduced by the imaging processes is spatially uniform, then the image may be described as a linear mapping of the object function. This mapping is of course at lower resolution, and the blurring is readily described as a convolution of the object function with a Point Spread Function.

Linear Imaging Systems

Image = object Point Spread Function + noise

OP=

O( x)P( x'x)dx

A convolution is a linear blurring. Every point in P is shifted, mapped and added to the output corresponding to the shape of O.

Linear Imaging Systems

Consider the simple model, a plane of sources I(x,y) mapped onto a plane of
detectors E(x,y).
The detectors measure photon intensity (energy) and do so in a linear fashion (if
twice the photon intensity impinges on the detector it returns a signal twice as
large).
Question: what would happen if the detectors saturated?

Linear Imaging Systems

There is a mapping from the source to the detector,

E( x, y) = S{I ( x, y)}
and the mapping is a linear functional so,

S{aI1( x, y) + bI 2 (x, y)} = S{aI1 (x, y)} + S{bI 2 (x, y)}

a and b are scalars.

Linear Imaging Systems

When we model the system as a linear functional then it is useful to introduce a

point source and to decompose the system into these. The point source can select
out every element of the input and follow how each element is mapped.
The benefit is that by focusing on how the points map we dont need to include the
object function in our analysis of the imaging process.

The Delta Function

The delta function allows a simple formal approach to the decomposition of the
image.
In 1 dimension,

0; x 0 ( x) = ; x = 0

The delta function is thus a singularity and has the following convenient properties:

(x)dx = 1

It is normalized and can select out a point of a second function,

f ( x) (x)dx = f (0)

provided f(x) is continuous in the neighborhood of x=0.

The Delta Function

The delta function also allows sampling at frequencies other than x=0,

0; x x0 ( x x 0 ) = ; x = x0

f ( x) (x x0 )dx = f (x0 )

So we can use the delta function to sample the object function anywhere in space.

The Delta Function

It is useful to introduce some more physical representations of the delta function:

lim

2 x 2

sin(x )
) lim sinc(x); sinc(x
= x

5

4 3 2

2 1

-2

-1

-2

-1 -1

Using these definitions you can show the various properties of the delta function.

The Delta Function

In[1]:=

g @ D : = Plot @ Exp @ Pi a ^ 2 x ^ 2 D, 8x , - 2 , 2 <, a_ a 8PlotRange -> All , PlotStyle -> Thickness @ 0.01 D, DisplayFunction -> Identity <D; Show @ Table @ @ H2 ^ n LD, 8n , 1 , 5 <D, g 4 DisplayFunction -> $DisplayFunction
2

In[2]:=

1.5

0.5

-2
Out[2]=

-1

Graphics

The Delta Function

In[3]:= In[4]:=

sinc @ D : = Sin @ D x ; x_ x s @ D : = Plot @ sinc @ x D Pi , 8x , - 2 , 2 <, a_ a a 8PlotRange -> All , PlotStyle -> Thickness DisplayFunction -> Identity <D; Show @ Table @ @ H2 ^ n LD, 8n , 1 , 5 <D, s 32 DisplayFunction -> $DisplayFunction
5 4 3 2 1

@ 0.01 D,

In[7]:=

-2

-1 -1

Out[7]=

Graphics

The Delta Function in 2D

The delta function is also straightforward to define in 2 dimensions,

where

0; r r0 ( r r0 ) = ; r = r0
2

r = xx + yy x and y are unit vectors.

r0 = x0 x + y0 y

The 2-dimensional delta function is therefore separable,

2 (r r0 ) = (x x0 ) ( y y0 )
We can also define the 2D delta function in cylindrical coordinates and this will be important for projection reconstruction, vida infra.

Linear Imaging Systems

Now that we have this useful construct of a delta function, let us return to our imaging system and decompose the plane of sources I(x,y). So any point in the source can be extracted as:

I ( x0 , y0 ) =

I ( x, y) (x x0 ) ( y y0 )dxdy

Linear Imaging Systems

We know the output of that point,

E0 (x, y) = S{I (x0 , y0 )}

Now notice that E is a continuous function over the detector plane, and so I have labeled the function by the position in the source plane.

Linear Imaging Systems

Now let us just explicitly write out the function:

E0 (x, y) = S I (x, y) (x x0 ) (y y0 )dxdy

Since S and the integrations are all linear we can change their order.

E0 (x, y) =

I (x, y)S{ (x x0 ) ( y y0 )}dxdy

Now we see that the mapping is described for each point in the object function, and that the object function itself simply provides a weight for that point. Of course it is essential that S be linear.

Instrument Response Function

Now we picture every point as being mapped onto the detector independently. The mapping is called the instrument response function (IRF).

E0 (x, y) =

I (x, y)S{ (x x0 ) (y y0 )}dxdy

E0 (x, y) =

I (x, y)IRF(x, y | x0 , y0 )dxdy

Instrument Response Function

The Instrument Response Function is a conditional mapping, the form of the map depends on the point that is being mapped.

IRF(x, y | x0 , y0 ) = S{ (x x0 ) (y y0 )}

This is often given the symbol h(r|r). Of course we want the entire output from the whole object function,

E( x, y) =

I(x, y)S{ ( x x0 ) (y y0 )}dxdydx0 dy0

E( x, y) =

I(x, y)IRF(x, y | x0 , y0 )dxdydx0 dy0

and so we need to know the IRF at all points.

Space Invariance

Now in addition to every point being mapped independently onto the detector, imaging that the form of the mapping does not vary over space (is independent of r0). Such a mapping is called isoplantic. For this case the instrument response function is not conditional.

IRF(x, y | x0 , y0 ) = PSF(x x0 , y y0 )
The Point Spread Function (PSF) is a spatially invariant approximation of the IRF.

Point Spread Function

So now in terms of the Point Spread Function we see that the image is a convolution of the object function and the Point Spread Function.

E( x, y) =

I (x, y)PSF(x x0 , y y0 )dx0 dy0

Here I have neglected noise, but in real systems we can not do that.

Magnification

A system that magnifies initially looks like it should not be linear, but the mapping can also be written as a convolution. The IRF of course must include the magnification (M), IRF(x, y | x0 , y0 ) = PSF(x Mx0 , y My0 ) and the image is,

E( x, y) =

I (x, y)PSF(x Mx0 , y My0 )dx0 dy0

Magnification

A simple change of variables,

x0 Mx0 y0 My0

Lets us rewrite the image as:

E( x, y) =

1 M

x y I ( , )PSF(x x0 , y y0 )dx0 dy0 M M

And the imaging system is again correctly described as a convolution. We also directly see that as the image is magnified its intensity decreases.

An Example, the Pin-hole Camera One of the most familiar imaging devices is a pin-hole camera.
a b

object function source

pin-hole

image

The object is magnified and inverted. Magnification = -b/a Known prior to della Porta ca. 1600.

The Pin-hole Camera, Magnification

arrow @ , y_ D : = If @ x_ Abs @ D <= 5 && Abs @ D <= 20 , 1 , 0 D + x y If @ > 20 && y <= 30 && Abs @ D < Abs @ - y D, 1 , 0 D; y x 30
Provides the mapping for an ideal pin-hole camera: a = source to pinhole (on axis) b= pinhole to screen (on axis) a = 10 ; b = 40 ; mag = - Hb a L; Idealarrow
DensityPlot 8PlotPoints

@ , y_ D : = H1 mag ^ 2 L * arrow @ mag , y mag D; x_ x

DensityPlot 8PlotPoints @ Idealarrow @ , y D, 8x , - 128 , 128 <, 8y , - 128 , 128 <, x -> 8128 , 128 <, Mesh -> False <D

@ arrow @ , y D, 8x , - 128 , 128 <, 8y , - 128 , 128 <, x -> 8128 , 128 <, Mesh -> False <D

100

100

50

50

-50

-50

-100

-100

-100

-50

50

100

-100

-50

50

100

DensityGraphics

DensityGraphics

An Example, the Pin-hole Camera 2 Notice, however, that the object function is also blurred due to the finite width of the pin-hole.
a b

object function source

pin-hole

image

The extent of blurring is to multiply each element of the source by the source magnification factor of (a+b)/a x diameter of the pin-hole.

Pin-hole Camera Blurring in 1D

Adds a finite diameter pin-hole but assumes that the response in spatially uniform (defines a convolution integral). r = 2; h@ DR_ D : = If @ Abs @ DRD <= r Ha + b L b , 1.0 , 0.0 D; Plot @ @ h DRD, 8DR, - 10 , 10 <, 8PlotRange
1 0.8 0.6 0.4 0.2

-> All <D

-10

-5

10

Graphics

1 0.75 0.5 0.25 0 -10 -5 0 5 10 -10 -5

10 5 0

In[8]:=

Plot3D

@ @ D * h @ - y D, 8x , - 10 , 10 <, 8y , - 10 , 10 <, 8Mesh -> False h y x

, PlotPoints

-> 8128 , 128 <<D

Pin-hole Camera Blurring in 1D

5 4 3 2 1

-10

-5

10

1 0.75 0.5 0.25 0 -10 -5 0 5 10 -10 -5

10 5 0

In[8]:=

Plot3D

@ @ D * h @ - y D, 8x , - 10 , 10 <, 8y , - 10 , 10 <, 8Mesh -> False h y x

, PlotPoints

-> 8128 , 128 <<D

Pin-hole Camera Blurring in 1D for a grid

0.3 0.6 0.2 0.4 0.1 0.2 -100 -50 50 100 -100 -50 50 100

Distortions of a Pin-hole Camera Even as simple a device as the pin-hole camera has distortions 1. Limited field of view due to the finite thickness of the screen.
a b

object function source

pin-hole

image

As the object becomes too large, the ray approaches the pin-hole too steeply to make it through.

Distortions of a Pin-hole Camera 2 Also, as the object moves off the center line, the shadow on the detector grows in area, (and the solid angle is decreased) so the image intensity is reduced.
a b

object function source

pin-hole

image

There are three effects, radial distance cos^2, oblique angle cos, and effective size of pinhole cos. Therefore cos^4. The general oblique angle effect goes as cos^3.

Distortions of a Pin-hole Camera Reduction in field of view due to oblique angle effect.
1 0.3 0.8 0.6 0.4 0.1 0.2 -100 -50 50 100

0.2

Cos^3
50 100

-100

-50

1 0.8 0.3 0.6 0.2 0.4 0.2

Cos^4

0.1 -100 -100 -50 50 100 -50

50

100

Contrast and Noise in a Pin-hole Camera For a screen of finite thickness some light will penetrate.
1 0.6204 0.8 0.62035 0.6 0.6203 0.4 0.62025 0.2 -100 -50 50 100

-100

-50

50

100

Object function

Full 2D analysis

250

250

200

200

150

150

Object function
100 100 50 50

0 0 50 100 150 200 250

0 0 50 100 150 200 250