Many different types of hyperspectral imaging systems are currently deployed in a range of applications including remote sensing, machine vision, surveillance and security and defense. However, despite their novel abilities, they have found a limited market due to their size and expense.
Now, researchers at PARC, a Xerox company, led by Dr. Alex Hegyi have developed a new novel hyperspectral imaging (HSI) system that could be added at minimal cost to existing camera sensors, such as those commonly found in cell phones and other consumer electronic devices.
According to Dr. Hegyi, a number of different technical approaches have previously been taken to create hyperspectral imagers. All of them, however, have one thing in common. They all acquire both spatial and spectral data across a range of contiguous spectral bands. The data are then combined to form a three-dimensional datacube that comprises a set of images layered on top of one another. Each of the layers comprises two spatial dimensions and one particular wavelength band, or spectral dimension.
One common technique used to capture hyperspectral images is to use a push broom scanner. Here, light from an image which impinges on an entrance slit is dispersed through a grating or a prism which separates the different wavelengths. The dispersed light is then captured by a sensor as a two dimensional image that contains both a spatial dimension and a spectral dimension. A second spatial dimension in a scene is generated by scanning the camera over the scene.
Such systems are highly effective at imaging scenes which move beneath the scanner. As they do, the push broom imager captures a complete spectrum of each point along one spatial line. Successive lines measured over time can then be used to form a complete hypercube. The drawback is that such systems suffer from very low light throughput, as the light from the scene is restricted by the slit. Another issue is that the dispersive optics can only disperse wavelength as a function of angle, and therefore the rays must travel some distance to be spatially dispersed at the detector. This means that the mechanical footprint of such instruments is large.
Another approach that has been widely employed to capture hyperspectral images is to use what is known as “staring” hyperspectral imagers. Staring hyperspectral imagers capture the spatial information of a scene instantaneously within a particular wavelength band. To do so, they employ wavelength-tunable filters placed in front of a standard CMOS or CCD imaging array.
Several types of filter can be used in such hyperspectral imagers. One technique, for example, uses acousto-optical tunable filters, but these are bulky and can require considerable power consumption. Another uses liquid crystal tunable filters – these rely on multiple stages of wave plates and liquid crystal phase shifters that work together enabling one wavelength band at a time to be selected. A third approach is to use MEMS-tunable Fabry-Pérot interferometer (FPI) filter structures that employ two reflective mirror surfaces. Here, the pass-band wavelength of the filter is tuned by adjusting the distance between the mirrors.
All these approaches have the advantage that one wavelength band at a time can be selected. But there are a number of disadvantages – one being that they have a fixed spectral resolution. For the liquid crystal (LC) tunable filter, the resolution is determined by the number of LC stages while for the FPI filter, for example, it is determined by the reflectivity of the mirrors. In addition, with all these filtering approaches, the spectral resolution is inversely proportional to the light throughput. So the narrower the spectral filter, the less light is transmitted through it, often necessitating long integration times. Finally, all these approaches have a maximum spectral bandwidth that is less than a factor of two between maximum and minimum wavelengths because of their limited free spectral range.
The last category of contemporary hyperspectral imaging systems use so-called “snapshot” filters to sample the full data-cube instantaneously without the need for tunable optical elements or mechanical devices. In such systems, a series of FPI filters are built onto standard CMOS sensors which enable the devices to acquire multiple spatial bands in parallel with each spectral band.
Hyperspectral imaging systems built using such devices are ideally suited to capturing scenes with high light levels and fast moving objects. However, they typically offer lower spatial resolution, as a consequence of the fact that spatial resolution has been traded off, in a fixed way, with spectral resolution. Also, as above, they have a limited free spectral range and thus a maximum spectral bandwidth less than a factor of two between minimum and maximum wavelengths.
The new system developed by Dr. Alex Hegyi and his team at PARC resolves many of the disadvantages of earlier designs, and has the added benefit of being relatively simple in its construction. Indeed, the system itself simply comprises a liquid crystal cell between a pair of polarizers that acts as a spectral encoder and a CMOS image sensor.
The effectiveness of the new concept has been demonstrated by building a prototype imager using a commercial 640x480 pixel CMOS camera.
The new design works in a manner not entirely dissimilar to a Fourier transform spectroscopy system, where different wavelengths of light pass though to a detector where the beam intensity is measured. The beam of light is then modified by moving one of the arms of an interferometer inside the system to yield a different combination of wavelengths. The resulting signal that is detected is called an interferogram, and has the property that every data point that comprises the signal has information about every frequency in it. The interferogram is then Fourier transformed by software into an actual spectrum.
However, rather than interfering light that travels across the two separate arms of an interferometer, Dr. Alex Hegyi’s new system employs a LC layer sandwiched between crossed polarizers. By synchronizing the drive of the LC with the camera’s image acquisition, the system performs interferometry between two polarizations of light that travel through the LC. The interferometric data are then analyzed to provide the spectral information.
More specifically, a polarizer in the system first polarizes incoming light in an incident polarization direction. The light then travels through the LC cell. By varying the voltage across the cell, the LC molecules change their orientation, and a variable optical path delay is created between horizontally and vertically polarized light. This path delay causes a wavelength-dependent phase shift between the two polarizations, leading to a wavelength-dependent change in the polarization state.
A second polarizer, or analyzer, changes this wavelength-dependent polarization state into a wavelength-dependent intensity by interfering the two rays, creating a temporal signal at each pixel that is the equivalent of an interferogram. A data set consisting of a series of images is then recorded by the CMOS imager as the path delay of the light through the LC cell is varied. Lastly, by Fourier transforming the data set with respect to the path delay, the hyperspectral datacube can be constructed.
According to Dr. Hegyi, the approach has some notable advantages. Because of the high optical throughput of the system, creating the hyperspectral image cube only takes around a second. The system has unlimited free spectral range, so its spectral bandwidth spans the sensitivity range of the CMOS imager. In addition, the spectral resolution, imaging speed, and spatial resolution can be traded off in software to optimally address a given application.
The researchers at PARC have already demonstrated the effectiveness of the concept, by integrating the LC variable retarder into a commercial 640x480 pixel CMOS camera. The prototype can acquire a full hyperspectral data-cube in 0.4 s, and is sensitive to light over a 400 nm to 1100 nm range with a dispersion-dependent spectral resolution of 450 cm−1 to 660 cm−1.
Dr. Hegyi believes that the simplicity, compactness and potential low cost of the device could enable a new class of hyperspectral imaging systems to be built. Indeed, because the complexity of the device has shifted from hardware to software, the sophistication of full spectral processing is now within reach of a far greater number of users.
Written by Dave Wilson, Senior Editor, Novus Light Technologies Today