Reassessing the Interpretation of Spectral Channels

In this interview, Frederick G. Haibach Ph.D., Principal Consultant at Confluent Sciences Consulting, Inc., talks to AZoM about how we should be interpreting spectral channels differently, and how that could affect the industry.

When did you first develop your interest in imaging?

In 2001, I noticed that most of the projects I was working on were as much about locating a feature as they were identifying the composition of the feature. At many different scales, we are interested in how an object, or scene of objects, are not homogenous. In traditional spectroscopy, we use sophisticated instruments to record the spectral response of a single spot. That spot may be large or small, compared to the feature of interest. If it is small, then we have little indication of the extent of the feature. In the extreme case, we might miss the feature entirely. If the spot size is large compared to the feature, then the surrounding area dilutes the signal. On the spectroscopy side, this means that the instrument has more stringent specifications than might be needed for the task.  The project then becomes more expensive and cumbersome than might be needed.

As an example, I worked on a project to identify silicone contamination on an injection molded foam part for automobiles. The automotive industry recognizes that silicones are an excellent release agent for molded parts, but silicones readily transfer to, and contaminate, other parts where they cannot be tolerated. Hence, the method needed to be non-contact and high contrast for silicones. During the injection molding process, excess silicone release agent is swept into a small region on the part being molded, resulting in a higher concentration in one portion of the part. The region of contamination would have been easily detectable using an inexpensive imaging instrument, but the technology available to us required multiple spot measurements.

How did working with Multivariate Optical Element’s frame your thinking?

Multivariate optical elements (MOEs) work well for enhancing contrast in a scene. The MOE is a type of dichroic filter with a designed transmittance and reflectance profile across the spectrum. Given a spectral signature of a feature of interest, a filter can be designed to create a transmittance and reflectance spectrum whose difference is related to that signature providing some of the advantages of imaging and spectroscopy. The result is more specific and sensitive than using a single wavelength. Using dual images in an imaging setup is challenging but it has been repeatedly demonstrated in an academic and commercial settings.

While there are advantages to using MOEs over spectrometers and single-wavelength imagers, the limitations are also great. The MOE is limited to its design criteria, and yet it is not easy to determine when that range is exceeded. That means that it does not fail gracefully. For example, image analysis of 3D objects is complicated by needing to evaluate both the difference and the sum of the transmittance and reflectance. Without this treatment, spurious results can occur due to illumination artifacts and surface texture.

Using spectrally structured light and off-the-shelf imagers, the advantages of MOEs can be realized along with the advantages of traditional spectrometry. Spectrally structured light is easily achieved using single-color sources like LEDs. When you think of the spectrally structured light as a set of patterns that provide contrast for spectral feature of interest, the result resembles the way that Fourier, wavelet, and other transforms highlight particular textures across an image. Much of this information can be encoded with just a few patterns or wavelengths that are judiciously chosen. In the silicone release agent example, a single wavelength infrared diode and a thermal camera would have sufficed. If rhodamine or fluorescein had been added to the silicone oil, a grayscale imager and an ultraviolet lamp would have detected the fluorescent tracer.

Where do you feel the fundamental differences lie in the interpretation of spectral channels?

Most imaging applications have focused on visible light in the manufacturing space, possibly because of personal experience. If we can see the defect with our eyes, then we should be able to measure it with a visible-light imager. There is a lot of specific information about the composition and structure of the material if we use information outside the range of visible light. As speed of measurement is important in manufacturing, fluorescence, reflection and absorption are the most commonly used. I have used instrumentation that ranges from the ultraviolet to the mid-infrared.

Ultraviolet light provides information about the surface, often less than a few hundred nanometers because materials are highly absorbing or scattering. The contrast in the ultraviolet between different materials can be quite dramatic. Some materials also fluoresce, which makes detection with silicon detectors much easier. Even faint fluorescence can be detected if the illumination is modulated.

The visible range provides the greatest sensitivity in silicon imagers. Even though our eyes provide only broad spectral information, mostly red, green and blue, there is more information present about the pigments and dyes that generate those colors. We can experience this directly by observing metamerism, the change in apparent color of an object under different “white-light” illumination. Additionally, polarization is useful, such as identifying the molecular order within a polymer part or strain in glass.

Near infrared (NIR) is a workhorse in food processing and manufacturing. Virtually all wheat sold on international markets has been characterized by a near infrared analyzer. A significant number of industries use NIR for detection of water and moisture. The advantages for NIR are the relative transparency of many materials allowing measurement of features millimeters under the surface and sensitivity to molecular structure. While metals do not have a molecular structure, the oxide coating may be measurable in certain circumstances. The spectral range is typically divided in two by available detector technologies. Short wave NIR can be used with silicon detectors for wavelengths up to 1100 nm.  Long wave NIR is used with InGaAs, InSb and thermopile detectors for wavelengths up to 5 μm. Long wave NIR has relatively high contrast, but imaging arrays have a high component cost and some manufacturers of sensor arrays have export restrictions.

Molecular information is the richest in the mid-infrared, but it is also the most technically challenging. Mid-infrared has been used to reveal composition and details of molecular structures, like the amount and type of crystallinity in polymers and inorganic materials. Other than lasers, light sources in the mid-infrared are dim. Thermopile and micro cantilever imaging detectors are relatively slow, but do not dominate the instrument cost. Since useful images don’t need that many pixels, mid-infrared imaging could become more commonplace.

How do these differences in spectral channels currently affect the manufacturing industry?

Insight into manufacturing processes is enhanced by information.  Spectral information provides unique insights about the properties that relate to structure and strength of materials and the analysis can be automated to provide tight, closed-loop control.

Manufacturing should not be slowed down by quality control. Silicon imaging detectors, like CMOS and CCD, are ideal for integration into the manufacturing line, and offer high sensitivity, spatial resolution and standoff distances from millimeters to meters.  When 100% inspection is not required, or when a quality attribute is difficult to access otherwise; slower methods can be considered using long wave NIR and mid-infrared.

Where do you believe changes can be made within the industry to improve spectral interpretation and performances?

Machine learning to recognize part outlines and 3-dimensional shapes is well developed with SDKs being offered by main-line camera manufacturers and third-party tools. Interpreting and designing spectral-based systems is still nascent. This may be because relating properties of the process to spectral signatures is usually not part of an engineer’s training. In addition, the algorithms used to recognize parts and textures in imaging are not well suited to the extra spectral dimensions. The converse is also true, that typical algorithms used for automated interpretation for spectra do not work efficiently on images. Bringing together the information from imaging and spectral data is possible, but the choice of how to acquire spectral data and processing still depends on domain experts.

After using spectral machinery, how does this affect process learning?

Pursuing reliability in both form and function, goes beyond shape and texture recognition to sub-surface and molecular properties. Optical methods also allow for rapid, non-destructive inspection. Improved inspection rates allow for dramatically finer process control with reduced risk for process over- and under-shoot. The defect analysis also improves understanding the variables that contribute to out-of-specification product. In manufacturing facilities where this type of process monitoring has been adopted, the improved process understanding has resulted in substantial cost savings.

About Frederick G. Haibach Ph.D.

Frederick G. Haibach, Ph.D. has been involved in process and industrial spectroscopy since graduating from the University of New Mexico in 2001. He has more than 15 successful projects in various areas of spectroscopy from the ultraviolet to the mid-infrared. As a consultant, he has three current projects in imaging spectroscopy.