AIA - Vision Online
Vision Resources

The Art of Lighting Science

by Kevin Harding - General Electric Corporate R&D Posted 04/28/2000

The tools of machine vision have seen many strides in the past few years.  Computer hardware has become faster and more powerful, sensors for machine vision have become more useful, and the software has proliferated for almost all applications.  One of the last areas seeing improvement has been the lighting and optics.  If the image is not good to begin with, processing can be a very difficult to impossible task.  Lensing issues are important to this process.  The aberrations and application problems must be dealt with if the image is to be the best that it can for processing.  But the tools to design a lens or find the best one for a particular application are fairly well defined. 

Lighting, on the other hand, has been referred to as something of a "black art."  The needs of lighting for machine vision are unlike the needs of almost any other lighting discipline.  What the camera and vision system need to see, and how they see it, is often much different from what the human observer sees.  It has been said that if bad optical design can make things appear other than they are, bad lighting can make features vanish entirely.  Too often the practitioner starts by just picking a tool, and trying to make it work.  If all you have is a hammer, everything looks like a nail.  By approaching lighting as a Science, we can begin to create an organized process for defining the problem, and collecting the information that will help identify the solution.

To the purpose of treating Lighting as a science, I suggest to consider four basic dimensions.

Application Requirements:
This dimension is the starting point of lighting determination.  In terms of machine vision, the application requirements specify the area of interest to be inspected.  It asks the question, "What am I looking for?"  The application requirement may be to detect some sort of feature on a part surface such as finding a hole in a deburring operation.  It may be to find flaws such as excess spatter in a welding operation or imperfections in glass products.  An application requirement may be to make a 2-D or 3-D measurement such as paint thickness on a car body panel or strain on a drilling tool.  To successfully use the tools of lighting as a science, you must first determine the application requirements.  These are the gateways to determining what lighting scheme to use.  A clear definition of the task is essential for whoever is going to do the application, least the wrong problem be solved, and the application is seen as a failure.

Problem Characterizers: 
The next dimension encountered in using lighting science as a tool is that of Problem Characterizers.  This dimension defines the conditions, requirements, and constraints of the inspection task.  It defines, "What you have to inspect?"  This dimension explores a field often neglected in the set-up of a lighting system: what is the "reality" of the inspection problem at hand?  To fully define the problem, all of the constraints must be identified.  This reality audit must address the light reflecting properties of the part to be inspected, the light level requirements or limitations imposed by the camera of environment, the need for a special character of light, the requirements of resolution or other image considerations, and finally the physical constraints and costs imposed at the installation site.  Let us review each of these sets of problem characterizers.
 
The optical performance of the part has a major impact on the ultimate performance of the machine vision system.  Ironically, the part characteristics are probably the most important in setting up a lighting system.  The following questions are usually posed on other optical elements of a system, but in the Problem Characterizer dimension, they are posed on the part.

The purpose of these characterizers is to cover every angle of how the part interacts with light, to the extent that it imposes some constraint on the ultimate design of the lighting system.  The constraint may be that the part is mirror-like (specular) or has many nooks and crannies (like an English muffin).  Although many part features may seem insignificant at first, there may be a technique which takes advantage of that very "insignificant" part feature that was overlooked before.
 
The process of analysis assumes that we are not starting with a "preconceived" notion on how the problem should be addressed, but rather leaves to door open to all possibilities until all the constraints of the problem have been defined.  Some basic part characteristic are as listed below.

Problem Characterizers:

I.       Part
     A. Surface finish
       1. Is the part highly specular and
               a. flat surfaced?
               b. curved in surface?
               c. irregular in surface?
       2. Is the part highly diffuse and
               a. flat surfaced?
               b. curved in surface?
       3. Is the part partially diffuse and
               a. directionally sensitive?
               b. does it have directionally uniform reflections?
       4. Are there mixed surfaces of specular, diffuse, and directional?
     
     B. Surface geometry
       1. Is the part flat?
       2. Is it gently curving?
       3. Does it contain sharp radii?
       4. Does it have mixed surfaces?
     C. Surface reflectance
       1. Is the part highly reflective?
       2. Is it poorly reflective?
       3. Does it have mixed areas of good and bad reflectance?
       4. Is it translucent?
       5. Is it transparent?
     D. Coloration
       1. Is the part monocolored and
              a. broad in spectral content (gray)?
              b. narrow or singular (i.e. red)?
       2. Are there color variations that are
              a. subtle?
              b. discrete?
       3. Is the part mixed with broad and discrete colors?

The problem characterizer dimension must also investigate the light level requirements in order to determine how much light the system needs, due to the camera sensitivity, or how much light is too much ("well, we finally got a bright enough light, but we have to get the data very quickly because the part begins to melt after a few seconds" - supply your own cartoon).  Things that cause low-light efficiency are large standoffs or poorly reflecting surfaces to name a few.  The human factor limit is exemplified by situations where humanly dangerous lighting such as high-power lasers or strobes is being used.   The following questions are aimed at identifying these constraints on the problem.

Problem Characterizers (continued):

II.      Light Levels
     A. Is the subject light sensitive?
     B. Is there low light efficiency?
     C. Is the area covered
          1. large?
          2. small and high in density?
          3. shaped?
     D. Is there low detector sensitivity?
     E. Is there a human factor limit of light?

We also must investigate questions about the character of the light in this dimension. The light character is a very important element in a vision system.  For example, if infrared readable codings are being inspected, the light source must produce enough infrared light for the system to detect the coding.  Pulsed lighting may come into play when the part under inspection is moving.  The duration and repetition must be calculated to handle the speed of the part.  The system must be able to "see" without blurring.  By answering the questions on light character requirements, the practitioner must consider under what conditions the inspection is to take place.

Problem Characterizers (continued):

III.      Light character requirements
     A.  Is a high light concentration needed?
     B. Are shadows required?
     C. Is uniform light critical?
     D. Is polarized light needed?
     E. Spectral content
         1. Is a broad color content desired?
         2. Are specific color bands needed?
         3. Is one narrow color band required?
     F. Is coherent light required?
     G. Pulse light
         1. Durations
             a. Are very short pulses needed?
             b. Are pulses over 10 milliseconds needed?
         2. Repetition rate
             a. Is a single pulse or low rate required?
             b. Is a high pulse rate needed?

The image requirements must also be addressed in the problem characterizer dimension. Determining what is required in the image outlines the task and defines the limitation of performance that can be expected from the viewing system.  For example, the resolution of the optical system must be defined (i.e., how small of a feature must be measured).  A simple shape identification task may be desired where high resolution is not needed.  However, high resolution is needed for accurately gauging to small tolerances.  As the resolution performance of the lens system is approached, the lens degrades the contrast of the image until the small dimensional changes are washed out of the image.

An initially low-contrast image produced by the lighting further degrades the limiting resolution of the viewing system.  In like manner, if subtle gray scale variations are important, harsh lighting producing near binary surface shadows would be undesirable.  Even the depth-of-field can be affected by the lighting to the extent the lighting of the features of interest changes or even vanishes of the range needed.  These issues lead to the next set of questions which characterize "what we have" in the particular application. 

Problem Characterizers (continued):

IV.      Image Requirements
     A. Is high resolution needed?
     B. Is silhouetting of a 
         1. flat subject desired?
         2. subject with rounded edges needed?
         3. partially transmitting subject needed?
     C. Is gray scale imaging required
         1. with edge enhancement?
         2. with absolute gray level control?
     D. Is a large depth of field required?

A constraint that can seriously alter the capabilities of a vision system is that of physical boundaries. The standoff obviously has an effect on the lensing, what focal length will be needed, and at what distance should the lens be designed to operate optimally. 

To demonstrate a little interchange between some of the Problem Characterizers, consider a setting with continuous vibrations.  The vibrations are so heavy that the part under inspection moves, thus answering yes to the vibration question.  We might move to the light character requirements and determine that pulsed lighting is a possible answer, but there are people working within the inspection area which imposes a human factors limit within the light level charaterizer.  It is rare, unfortunately, that there can be much change in the manufacturing machinery to accommodate a more effective machine vision implementation, so these constraints must be considered in the lighting design.  Typical physical constraints to consider are listed here. 

Problem Characterizers (continued):

V.      Physical Constraints
     A. Standoff
         1. Is there large standoff?
         2. Is there short standoff?
         3. Is there a fixed distance?
     B. Work volume
         1. Open volume?
         2. Small access with large volume?
         3. Small access with small volume?
         4. Is it unique in shape?
     C. Vibrations
         1. Are there continuous vibrations?
         2. Are there shock vibrations?

Economic considerations must always be addressed when designing machine vision systems.  You can't do much if you don't have the money!  The cost of a machine vision system extends beyond just the equipment and development.  The problem characterizer exercise must answer key economic issues.

Questions such as:

Very often, people begin machine vision ventures without correct funding information.  Often, a little extra money spent up-front on the lighting to provide a better image, before the major commitment in software development is undertaken, can serve as a major savings later in the project.

Techniques/Tools:
The third dimension of lighting science explores the actual techniques and tools for machine vision tasks.  This is the "meat" of lighting science.  Numerous techniques are available according to how the problem characterizers are answered. A list of potential techniques and tools may look something like this:

Techniques:
1. polarized lighting
2. strobe lighting
3. grid enhancement contouring
4. multiple direction lighting

Tools: 
1. polarizing fitters, wave plates, etc.
2. strobe light
3. grating or grid projection system
4. ring light

A bad practice, which is unfortunately common, is to begin with a technique and find an application for it.  This method puts the practitioner at a great risk of losing money.  Starting with a technique already in mind may cause a compromise in the goals of a real problem just to make the technique fit.  The logical flow of machine vision system design is to begin with some set of requirements, characterize what you're looking at, and then converge upon a technique with many references to back you up.

Resources/References: 
The final dimension of lighting science is the wealth of published information describing what others have already done.  A technique is useless if you can't find the tools to make it work. One reference might, in turn, suggest further literature references
to aid in the overall understanding of the technique.  Further literature may even be the answer to finding the specific technique that's right for you.  The user should be led to an understanding of the state of the suggested technique in other fields.  Locating the manufacturer of certain tools may be easy, but first you must know if the tools are even available?  This dimension is the final chapter to lighting science.

Applications:
The user and the database searches for appropriate techniques and references answer the questions listed previously.  To give a feel for how lighting science functions, we will describe a number of examples.

Suppose Company A wants to measure size characteristics of a machine part (i.e., how wide it is, how long, how thick, etc.).  The part is actually a turned shaft.  Its surface is very specular.  The thickness varies because the shaft is tapered.  The part is to be inspected on a conveyor prior to packaging.  The conveyor will stop for inspection, if necessary.  The designated area in which inspection will take place is located in a big warehouse-type room.  There are support structures for machinery that extend over the conveyor and prevent a camera from coming closer than 8 feet to the parts.  All of these constraints define the application requirements.  The next step is to characterize our problem. We characterize the turned shaft as follows:

We chose low light efficiency because we are far away from the part (8 ft.).  Also, the large area is well lit and, without shrouding, this extra light could affect the system in the form of noise.  These types of possibilities are brought out when we characterize a problem. A listing of possible techniques might include:

In this case, polarized light might be a good one to consider. Polarized light is used to reduce the glinting produced by specular surfaces.  We might also consider the use of coherent light or other monochromatic (single color) light and a color filter to separate the designed lighting from the background because of the high intensity room lighting.  We must make sure that we choose our lensing to cover the distance between the system and the part.

What if the original constraints were changed?  Let's say that instead of the part being entirely specular, that part of it is painted.  Size characteristics are still the desired inspection.  The painted surface is directionally diffuse and not specular.  So, you still get glints, but they are not specular glints.  Review of articles on polarized lighting would tell you that you couldn't use polarized lighting for the painted area because the reflection from a diffuse surface does not maintain the polarization of the incident light.  The suggested solution may be to illuminate at one angle and view at a greater angle so that you "miss" the directional reflection. 

Now let's say that we are able to shroud the part and the inspection station together.  Now you don't have to use laser light, because the outside noise has been reduced.  Furthermore, what if Company A decides they want to check the surface quality of the shaft.  With the characterizers that we specified, a low-angle directional light might bring out these surface features.

Company B manufactures glass containers for beverages, pharmaceutical, and many other industries.  They want to inspect their containers for defects.  The defects are either cosmetic (displeasing to the eye) or functional (affect the performance of the container).  Some defect examples provided by Company B include bubbles, cracks, and stones.  Bubbles form when the glass doesn't anneal correctly and oxygen gets trapped in the material.  These defects are usually cosmetic unless they are very large.  Stones are small pieces of foreign material that get caught in the molten glass.  Particles like dirt and furnace chips make up stones.  These defects are usually cosmetic, also.  Cracks are obviously functional defects.  They form when glass containers bump into each other or when they are handled too roughly.  Improper annealing can also produce weak spots in the glass which are more susceptible to cracking.  The inspection system should be able to inspect a number of different containers.  The glass is transparent and the defects need only be found and not necessarily measured.  With these application requirements in mind, we characterize the part as follows:

There is low light efficiency in a reflection-based system because most of the light passes through the glass and does not get reflected.  With these characterizers in mind, a review of the literature might suggest the following:

A description for dark field illumination would explain that light (diffuse) is directed to the glass and the sensor is positioned so that it will not see the specular reflection.  What the sensor sees is a relatively dark field over the glass.  If any defects are present, they appear as bright spots because they scatter the light toward the camera.  Cracks are illuminated on the principle of total internal reflection.  We are also advised to experiment with the angular spectrum of the illumination light to make defects more visible.  You can tailor your light to the high- or low-angle portions of the scattering spectrum to improve the signal to noise ratio.

What if the constraints changed and Company B wanted to know the size of the defects?  The literature may now suggest that for such exact measurements you would have to use a transmission-based (illuminate from behind).  Or, what if Company B wants to inspect translucent containers, also?  A translucence implies that you can see some light through the material, but that light is scattered widely.  Images are difficult or even impossible to perceive correctly through a translucent material.  You may have to rely entirely on reflectance techniques for such a situation, depending on the degree of translucence.

CONCLUSIONS:   
It doesn't matter how good or bad a scene looks to you, it's how it looks to the machine vision system that matters.  A step-by-step characterization of the application, part, and constraints determines how the scene is presented to the sensor.  Machine vision has grown enormously.  However, it has always been hampered by the lack of development of its first element: lighting.   This paper reports on a beginning of tools to guide the lighting in machine vision applications.  With the approach of lighting as a science, the practitioner can now begin to take steps forward rather than stumbling in the dark.  Through the use of the logical dimensions of application requirements, problem characterizers, techniques/tools, and resources/references, lighting is made easier and faster.  By making use of the various sciences involved in past lighting research efforts, we can now work toward a "brighter" (or at least a well-lit) tomorrow.


About the Author:  Kevin Harding is a Senior Engineer GE Research and Development Center.  He has over 23 years of experience working in the field of optical-based measurement and inspection.  Mr. Harding was awarded the Eli Whitney International award by SME as well as leadership awards from SME, the Automated Imaging Association, and the Engineering Society of Detroit.  He has been a frequent author, chair, and instructor at many events for SPIE, LIA, SME, AIA, and others.   

Top of Page Email This Page Print This Page