ASM Metals HandBook Vol. 17 - Nondestructive Evaluation and Quality Control

Подождите немного. Документ загружается.

(d)

Typical performance

Reference cited in this section

11.

J.J. Dongarra, "Performance of Various Computers Using Standard Linear Equations Software in

a Fortran

Environment," Technical Memorandum 23, Argonne National Laboratory, 1989

Digital Image Enhancement

T.N. Claytor and M.H. Jones, Los Alamos National Laboratory

Image Capture and Acquisition System

An important part of image enhancement is the system that captures or acquires the data. In many cases, an image will be

recorded on film to be digitized, or the data will come directly from an inherently digital system, such as real-time

radiography, thermographic, or ultrasonic imaging system. The quality and dynamic range of the image will influence

subsequent processing. It is usually desirable to have interscene dynamic ranges greater than 6 bits (64 levels). The eye

has a dynamic range of about 10

(photopic and scotopic combined), an interscene dynamic range of about 150, and a

brightness discrimination of 2% (50 levels) (Ref 3). The schematic of a typical display system is shown in Fig. 2. Each

pixel can represent a gray level or color, and if the pixel is composed of 8 bits, 256 gray shades or colors can be

displayed. Because the eye has an approximate dynamic range of 150, a 256-level display is adequate for gray-level

viewing. For computational purposes, for the display of color, and for the ability to fully digitize film, it is sometimes

desirable to have a 12 (or more) bit pixel, which can be compressed into 8 bits for display. The display size indicated in

Fig. 2 is N × N; for most systems, N should be at least 512, with 1024 preferred.

Fig. 2 Concept of image display (a) and representation (b) in computer memory

Table 2 lists several types of digitizers that are currently available. Although charge-coupled device (CCD) cameras and

light tables are the standard for the efficient input of medium-resolution optical data, there are other alternatives better

suited to specific applications. The vidicons offer the highest frame rates in formats of 1024

and more, the line scanners

have excellent resolution at moderate prices, and the micro-densitometers are the most precise with regard to photometric

and geometric quantities. Most microdensitometers will digitize to 12 bits, which limits their dynamic range to about 3.6

in film density. As noted in Table 2, the standard scanning microdensitometer (row one) will take longer (up to several

hours) to scan film if the film density is over 2. The laser scanners are faster (depending on the make) because of the more

intense laser light source and the fact that one axis is scanned with the light beam rather than mechanically.

Table 2 Typical characteristics of image digitization devices

Face plate

illumination

Type Image

quality

Pixel size or

pixel number

resolution

Acquisition

time per

1024

by 8-bit

image

Interscene

dynamic

range

Image

area, mm

× mm

(in. × in.)

ftc

Microdensitometer Excellent

5-75 m (200

in.-0.0030 in.)

(a)

15 min

4096

500 × 500

(20 × 20)

. . .

Point source

and detector

Laser scanner Excellent

35 or 75 m

(0.0014 or 0.0030

in.)

(a)

1 min

4096

430 × 350

(17 × 14)

. . .

Linear CCD array Good

4096

500 ms

>1500 . . . . . .

. . .

Linear detector

Photodiode array Good

1024

5 s

1000 . . . . . .

. . .

CCD Fair 1320 × 1035

140 ms

500 . . . 1.5

0.14

Solid-state

cameras

CID Fair 776 × 512 33 ms

(b)

200 . . . 0.1

0.009

Vidicon (Sb

) Fair >1500 33 ms, lag =

20%

80-200 . . . 5

0.5

Newvicon (ZnSe) Fair 800 33 ms, lag =

10-20%

50-200 . . . 0.5

0.05

Pasecon/Chalnicon

(CdSe)

Fair 1600 33 ms, lag = 5-

10%

30-60 . . . 1

0.09

Saticon (Se+Te+As) Fair 1200 33 ms, lag =

50-160 . . . 6

0.6

Plumbicon (PbO) Fair 1200 33 ms, lag =

80-160 . . . 2.5

0.23

Tube type

cameras

(vidicon family)

SIT (Silicon) Fair 750 33 ms, lag =

60 . . . 0.01

0.0009

(a)

Aperture.

(b)

776 × 512

Linear CCD or photodiode arrays are good choices for film digitization and can have even better dynamic ranges than

those listed in Table 2. When these devices are cooled, each 7 °C (13 °F) reduction in temperature reduces the root mean

square noise by a factor of two. The charge-coupled device and charge injection device are capable of good dynamic

range and fair resolution, but have certain artifacts (Ref 12). Again, if the devices are cooled, the noise floor is reduced,

and they can be integrated for long periods to enhance the dynamic range and sensitivity. Cooled CCDs are available that

rival the low light sensitivity of silicon-intensified targets (SITs), but not the frame speed.

In general, the frame rate, dynamic range, and resolution are all interrelated. The interscene dynamic range is listed as the

maximum achievable for the microdensitometers and as the dynamic range that can be achieved at the given frame rate

for the other devices. The faceplate illumination given for the tubes and the solid-state detectors assumes mid-level

illumination (halfway between saturation and preamplifier noise) (Ref 13). This may vary among tubes and generic types

by a factor of three to four (Ref 14). The interscene dynamic range will also vary greatly, but can be maximized by the

proper selection of a tube such that a dynamic range of 200 can be achieved at 33 ms/field with a high resolution (>1000

lines).

The quoted resolution for the tubes is at a modulation transfer function (MFT) of 5%, which means that the contrast

between a black line and a white line is only 5% at the stated resolution. This, of course, is measured at optimum

illumination and at the center of the tube image field. Under other conditions, the resolution will be less. For comparison,

a 1024

CCD camera may have an MTF of 5% at 750 lines. The lags quoted for the tubes are typical for the particular

type at 3 TV fields or 50 ms. Tube cameras are primarily used in radiation environments or in specialized applications,

such as high-resolution real-time radiography, in which the frame rate is higher than that achievable by current CCD

designs. Charge-coupled device camera design is rapidly evolving for high-resolution scientific use and can be expected

to improve with regard to real-time frame rates (Ref 12, 15).

It should be noted that a 355 × 432 mm (14 × 17 in.) film digitized at a resolution of 50 m (0.002 in.) with 12-bit

accuracy will consume a 92-Mbyte file. Just writing or reading this file to or from a hard disk could take up to 8 min

(optical disks take even longer). Even with high-density optical disks and data compression, the digital storing of high-

resolution radiographs represents a formidable problem.

References cited in this section

3. R.C. Gonzalez and P. Wintz, Digital Image Processing, Addison-Wesley, 1977

12.

J.R. Janesick, T. Elliott, S. Collins, M.M. Blouke, and J. Freeman, Scientific Charge Coupled Devices,

Opt.

Eng., Vol 26 (No. 8), 1987, p 692-714

13.

I.P. Csorba, Image Tubes, Howard W. Sams & Co., 1985

14.

G.I. Yates, S.A. Jaramillo, V.H. Holmes, and J.P. Black, "Characterization of New FPS Vidicons for

Scientific Imaging Applications," LA-11035-MS, US-37, Los Alamos National Laboratory, 1988

15.

L.E. Rovich, Imaging Processes and Materials, Van Nostrand Reinhold, 1989

Digital Image Enhancement

T.N. Claytor and M.H. Jones, Los Alamos National Laboratory

Image Processing

Prior to image processing, it may be necessary to perform some type of preprocessing on the data. Most often, an image

file will need to be converted into the workstation standard from some other type of format; for example, the files may be

8-, 16-, or 24-bit uncompressed format with a variable-length header and need to be converted to 8-bit format with no

header. Typically, a problem arises when a file must be read that is in an unknown size, a compressed format, or has a

limited color or nonstandard palette.

Other preprocessing functions operate on the raw data in some manner before inputting it for image processing. For

example, the system should have the capability to average, filter, and acquire full frames and control the scanning

parameters on the digitizer so that the noise level can be minimized. In other cases, more complicated operations are

called for, such as tomographic reconstruction, synthetic aperture operations, or an FFT. Shown in Fig. 1(a) and (b) in the

article "Use of Color for NDE" in this Volume are tomographs taken before and after oversampling by a factor of four

and Wiener filtering of the input data set to correct for the point spread function (Ref 16). The preprocessing operation

greatly increases the smoothness of the lines and reduces the artifacts.

Reference cited in this section

16.

K. Thompson, private communication, Sandia National Laboratories, 1988

Digital Image Enhancement

T.N. Claytor and M.H. Jones, Los Alamos National Laboratory

Image Enhancement

There are three major scientific applications for image processing:

• Enhancement of an image to facilitate viewing

• Manipulation and restoration of an image

• Measurement and separation of features

These functions are listed in Table 3. Many of the functions have a dual purpose and, as will be shown, are usually

combined to form other functions.

Table 3 Image-processing software algorithms useful for NDE applications

Image enhancement Image

operations

Information extraction

Contrast stretching Scaling

Image statistics

Histogram equalization Translation

Point, line, angle perimeter, area, measurement

Contouring Rotating

FFT transformation, one and two dimensional

Thresholding Registration

Correlation

Composite image building Warping

Edge detection

Palette operations Combining

Deblurring

Color model selection Filtering

Motion restoration

True color representation

Thickening, thinning

Noise cleaning

Trend removal

Pattern recognition

Contrast Stretching and Histogram Equalization. The concept of contrast stretching is shown in Fig. 3(a). The

basic operation is given by the pixel value transform:

s = T(r)

s = T(r) = ar + b Linear

T(r) = a log(r) + b Logarithmic

(Eq 1)

where r is the original pixel value and s is the transformed pixel value.

Fig. 3

Concept of histogram stretching and equalization. (a) Histogram is stretched with a linear

transformation. (b) Histogram is equalized such that the probability density is constant.

The linear stretch and offset method is the most commonly applied, while transforms of the nonlinear type can be used to

convert film density to integrated dose or to linearize the film transfer function to account for base density. Although the

manual contrast stretch is often used, a more automated contrast stretch is available with an operation known as histogram

equalization. The general form is:

(Eq 2)

A simple linear equalization example is shown in Fig. 3(b). Occasionally, it may be desirable to use the cubic or

logarithmic equalization. The cubic emphasizes the lower values, making the image darker, while the exponential makes

the image much brighter. Other types of transfer functions are often used to remove nonlinear response in the imaging or

camera system. For isolated defects, a manual stretch usually gives good results. On textured materials, however,

histogram equalization often works well. An example of a linear histogram equalization of a low-contrast ultrasonic

image of a bond line in an explosively bonded steel-to-aluminum plate is shown in Fig. 2 in the article "Use of Color for

NDE" in this Volume.

Contouring and Thresholding. Contouring and thresholding are often combined with palette operations to show

gradations in intensity or to highlight defects. Color should be used to detect gradients in intensity because the eye is

sensitive to many more colors (approximately 10

shades) than gray levels. Thresholding can be done with a binary

picture (black and white), a color scale, or a combination of gray and color. The concept of thresholding has particular

application in the generation of a mask for future image-processing operations or the removal of the background.

Contouring is used to change only those values between two pixel levels to a certain value. It can also be used in outlining

high-definition features before further processing.

Composite Image Building. The ability to display composite images is useful in comparing the difference between

images produced by different modalities, such as a radiograph and an ultrasonic or NMR image. Because image quality

suffers greatly when displayed at resolutions of less than 512

it is imperative to have at least a 1024

display or to use

two monitors if detailed comparisons are to be made. Figure 4 illustrates how x-ray radiographs and neutron radiographs

can be displayed together and combined to enhance voids in a composite material. As shown in Fig. 4(c), the images have

simply been added to enhance the detection of voids. The absorption of neutrons in the epoxy is higher than in the

alumina and the reverse is true for the absorption of x-rays. However, the voids do not absorb in either case, and the

contrast is therefore improved if the images are added.

Fig. 4 Voids detected in three 60 × 60 × 10 mm (2.4 × 2.4 × 1 in.) alumina-

filled epoxy tiles using three

different imaging techniques. (a) Conventional radiography. (b) Neutron radiography. (c) Combined x-

ray and

neutr

on radiographic image. Building a composite image of both modalities indicates that there are voids in the

sample. The addition of the two images enhances the voids because of the differential absorption in the matrix

between the x-ray and neutron images.

Palette Control. Color can be displayed as a true color map or as a pseudocolor representation. In a pseudocolor

representation, the digital data values are mapped to any color that can be produced by the combination of the red, green,

and blue (RGB) values specified by a look-up table (LUT). A typical 8-bit pseudocolor system is shown in Fig. 5(a). With

only an 8-bit-deep pixel as shown in Fig. 2, a type of true color is possible; however, each color is represented only by

eight or fewer color levels (Fig. 5b). With 8-bits per color (as in Fig. 5c), the colors can be shaded continuously and are

suitable for three-dimensional displays in which depth cueing is indicated by decreasing the luminance. Eight-bit pixels

with 24-bit LUTs can produce good color scenes by adjusting the LUT to include only those colors that appear in a scene.

However, the palette is then image specific.

Fig. 5 Concept of a look-up table for three types of color displays. (a) 8-bit pseudocolor representation. (b) 8-

bit true color map. (c) 24-bit true color map. The LUT is controlled by the palette, which specifies what color

composition (R,G,B) will be assigned to one of 256 levels, as in (a). The digital-to-

analog converter (DAC)

converts the digital signal to RGB for input to the monitor. Shown in (b) is a way to display true color with an 8-

bit pixel. Better true color can be obtained by using a 24-

bit pixel, with each byte (8 bits) assigned to a specific

primary color, as in (c).

Several types of pseudocolor and linear gray-scale palettes are shown in Fig. 4 in the article "Use of Color for NDE" in

this Volume. The spectrum palette (or inverse spectrum) is preferred for range of color, but is not suitable for displaying

rapid changes in pixel level. Other palettes, such as the complementary palettes shown in Fig. 4(d) and (e), can be used

with good results on figures that have gradual changes in gray level. A typical threshold palette is shown in Fig. 4(b). All

values below 192 will appear as a gray scale, while those equal to or above 192 will appear as red. The gray scale is still

used to display lower-intensity data so that the operator can still detect unusual, yet not rejectable, anomalies. In contrast

stretching, when further work on the image is anticipated, the palette (that is, LUT values) is altered, leaving the pixel

values intact.

Color Models and the Use of Color. There are two main color models: the red, green, blue (RGB) model and the

hue, saturation, luminance (HSL) model. High-resolution imaging systems use the RGB model, while the National

Television Systems Committee has adopted the HSL model for broadcast television. The colors in the RGB model are

simply an additive mixture of the Commission Internationale de l'Éclairage monochromatic primary colors (red = 700 nm,

or 7000 Å; green = 546.1 nm, or 5461 Å; and blue = 435.8 nm, or 4358 Å). When equal intensities of these colors are

combined, a nearly white light results. In the HSL model, the primaries are transformed such that the hue value represents

the color (0 to 360°; red = 0, 360; green = 120; and blue = 240), the saturation is the color intensity (values 0 to 1), and

the luminance is the amount of brightness (0 to 1). If the saturation is 0 and the luminance is 1, the color will be white.

The advantage of the HSL model is that a single hue can be manipulated in intensity more easily than with the RGB

model.

Although both color and black-and-white images contain the same information, there are features that are much more

obvious in color plots than in corresponding black-and-white plots, and vice versa. In particular, a black-and-white

representation is appropriate when there is a high spatial frequency inherent in the image; in such a case, the mind-eye

combination interprets this as a texture. On the other hand, color is preferred when there are a few isolated objects or low

spatial frequency. This makes it easier for the image interpreter to discern small changes in image density through the use

of contrast enhancement. Black-and-white images are shown in Fig. 6 and 7, and the color versions are shown in Fig. 6(c)

and (d) in the article "Use of Color for NDE" in this Volume.

Fig. 6 Ultrasonic image of a hot isostatic pressed tungsten

plate prepared from powder (99% dense). The

black-and-

white image shows small changes in density that are not easily discernible. The density changes are

enhanced by the use of color, as shown in Fig. 6(c) of the article "Use of Color for NDE"

in this Volume. The ring

is caused by a transition from one thickness to another (the outer thickness was about 50% of the inner circle).

Fig. 7 Tungsten plate similar to the one shown in Fig. 6

, except that this part was fabricated from a plate that

was rolled rather than hot isostatic pressed. The black-and-

white image shows a texture imparted to the

material due to the inclusion of 2% porosity. The color version of the plate, Fig. 6(d) in the article

"Use of Color

for NDE" in t

his Volume, shows how difficult the texture is to interpret in a color representation using a

spectrum palette.

Digital Image Enhancement

T.N. Claytor and M.H. Jones, Los Alamos National Laboratory

Image Operations

Listed in the second column of Table 3 are some of the more important image operations that are routinely used to

manipulate images.

Geometric Processes. The scaling, rotation, translation, warping, and registration of an image are classified as

geometric processes.

The scaling of an image is an important function because it is often used before images are combined. Scaling is used to

magnify or shrink the images permanently to fit printer sizes or simply to view areas closely.

Rotation and translation are important mainly if two images need to be registered closely and combined or

compared. The operations are described below. The scaling operation can involve interpolation, or it can merely involve a

duplication of pixels. In the case of integer scaling (×2, ×3), the pixels are replicated in the x and y by the integer. For

noninteger scaling (such as magnification by 1.5), every other pixel is replicated. Image interpolation and noninteger

scaling should be avoided unless the image is subsequently filtered or unless interpolation and scaling are the last steps in

a processing chain before printing. The printer will often perform interpolation and dithering (adding a small random

number to the value) to obtain an image with a smoother appearance: