1. Introduction

Terahertz (THz) applications, typically between 0.1-10 THz (3000-30 $\mu$m), can be found in a range of areas, which include medical, security, communications, astronomy, food, agriculture and industrial applications such as quality control, sensing, monitoring and process control [1–3]. Unlike infrared radiation, THz radiation is highly transmissible through a range of non-conducting materials and allows imaging of concealed metal and non-metal objects [2]. Due to its low photon energy ($\sim$4 meV at 1 THz), THz radiation is non-invasive and non-ionising, and thus attractive for medical imaging and biological samples [4]. It provides spectroscopic responses to many materials, and for imaging, offers increased spatial resolution and higher contrast in comparison with currently available microwave and mm-wave systems.

These properties offer potential in a suite of applications, including the field of threat detection [5]. Many materials of interest for security applications, including drugs, explosive-related compounds (ERC) and explosives, such as PETN, TNT, RDX and HMX, can be measured and identified with THz spectroscopy [6–8] as well as modelled using density functional theory (DFT) [9,10]. This, along with the ability to penetrate many packaging materials and clothing [8], makes THz-based techniques attractive for use in person-borne explosive detection systems [11]. However, most studies to date have been laboratory-based, and transitioning the technology for in-field use is challenging [8].

A hyperspectral imaging (HSI) system can be formed by combining imaging and spectroscopy. Hyperspectral systems, where a large number of spectral channels effectively form a continuous spectrum, as opposed to multispectral systems comprising a few discrete spectral bands [12], generally offer more accurate identification of materials [13,14]. The geometry of a HSI system can be transmission [15,16] or reflection [17,18], where the 2D image is usually derived by moving the sample [15,16], moving the equipment [17] or by full-field imaging [18].

For threat detection applications, where materials are often present in bulk, a transmission measurement geometry is not always practical because the materials are too absorptive, resulting in little or no THz signal, plus if the object is on the ground, access to the other side is impossible without disturbing it. Thus, reflection is the most realistic approach, given the signals are largely independent of material thickness and measurements can be performed from a single side. Furthermore, reflection in combination with a phase-sensitive technique such as time-domain spectroscopy (TDS), allows 3-dimensional (3D) surface relief mapping of objects or their sub-layers [19–21] to determine their shape or structure, while still maintaining information about the reflectance and sometimes the spectra of the objects.

To allow detection in such scenarios, it is desirable to steer the THz radiation across an object, rather than move the equipment or move the object itself. A challenge arises from measuring a large area and depth range to obtain an image of an object, while simultaneously obtaining its spectral information.

Beam steering has been demonstrated, for example, using galvanoscanner systems [22,23] and a scan lens to rapidly steer a THz beam across an object. Common two-mirror galvanometers lead to distortion of the images, requiring software corrections [22] or compensation with anamorphic optics [23]. To minimise such issues, a single mirror on a gimbal can be used instead [24]. These approaches have been limited to pure imaging, incorporating neither spectroscopy nor HSI to identify sample materials. Additionally, these systems had relatively close measurement distances ($\sim$100 mm) and small depth ranges ($\sim$10 mm), which are insufficient for many in-field applications, requiring at least an order-of-magnitude greater. These demonstrations used free-space optics and/or bulky lasers, which are, again, not well suited to external, portable situations.

We address some of the above challenges to develop a fiber-coupled benchtop-sized THz hyperspectral imaging system as a step towards compact, portable field-deployable systems for security and industrial applications. This requires a large standoff distance to reach around obstacles, plus a good depth range and field of view to cope with bulky objects. To achieve this, the system is designed for a measurement volume of 150 mm diameter by 255 mm depth, with a standoff range of 1 m. To attain this, the system uses beam-scanning in a reflection geometry, based on a single, fast-steerable mirror in conjunction with a custom lens to scan the THz beam across a fixed object. This allows rapid scanning of objects, while combining spectroscopy and 3D mapping. Hyperspectral images are obtained by taking spectroscopic measurements at each point on an object and creating a THz image from the spatio-spectral information. Within the same measurement, high-resolution 3D maps are obtained by utilising the temporal information of the reflected THz pulses.

The spectral analysis capability was investigated using a saccharide and two acids, namely $\alpha$-lactose monohydrate, L-tartaric acid and 4-aminobenzoic acid (4-ABA). These materials exhibit a number of characteristic spectral features between 0.5-1.6 THz and are convenient simulants for a range of explosives such as PETN, RDX and HMX, respectively [25,26].

2. System design

The THz beam-scanning reflection system is based on a THz time-domain spectrometer (Menlo Systems TERA K15) with an integrated optical delay line offering up to a 1700 ps scan range. With reference to Fig. 1, fiber-coupled outputs drive a photoconductive antenna (PCA) emitter, while a PCA detector is simultaneously gated. A convex TPX lens (f = 50 mm) collimates the emitted THz beam with a beam diameter of 16 mm (FWHM). A 6 mm thick high-resistivity silicon beamsplitter (BS) divides the beam into a transmitted and a reflected beam with a uniform $\sim$54/46 (Tx/Rx) ratio over the applied frequency range. The transmitted beam can be utilised to monitor the signal strength of the system in real time, while the reflected beam is directed towards the scanning system.

 

Fig. 1. Schematic (left) and photograph (right) of the THz scanning-beam reflection system. The enclosure and tube system allow for purging with dry nitrogen to conduct low-humidity measurements.

Download Full Size | PDF

To overcome the aforementioned issues with two-mirror systems, the scanner was designed with a single 75 mm diameter gold-coated mirror mounted on a motorised two-axis gimbal (Newmark GM-6) controlled by a stepper motor driver (Newmark NSC-A2L) and a custom telecentric f-theta scan lens. The gimbal scans the THz beam across the scan lens, which consequently scans a focused beam across the object. THz signals reflected by the object are recollimated by the scan lens and directed by the scan mirror back through the beamsplitter and a second focusing TPX lens (f = 50 mm) onto the PCA detector. The 2.5 m air path between the THz emitter and detector was compensated by an additional 1.7 m length of optical fiber connected to the PCA detector to synchronise the time gating. The optical fiber has inverse dispersion to maintain the femtosecond pulse widths and preserve the THz bandwidth.

The optical system was placed into a custom-made acrylic enclosure with a connected tube system to optionally conduct reflection measurements at low humidities by purging with dry nitrogen gas.

Software was developed in LabVIEW to control the beam and synchronise the THz system with the two-axis gimbal to generate fully-automated HSI measurements. The software records the time-domain information at each raster scan position, before stepping to the next position. THz signals are registered and updated at a maximum of 20 Hz, while the time-domain information of the signal is used to obtain the spectral information using the Fourier transform. The software also allows the user to select any region of the raster scan image and display the spectrum of the material at that location.

A reference signal forms part of most spectroscopic measurement processes. In recent years there has been research into reference-less reflection measurements [12]. Examples include time-domain measurements for layer thickness analysis [21], or time-domain measurements for identification of materials through empirical classification [27]. These usually assume some knowledge of the materials and/or use a pseudo-reference in the form of a substrate or cover material.

A good reference is important for producing high-quality object spectra prior to analysis, by removing common spectral features due to the instrument and the environment [14]. For this reason, we use a physical reference surface in the form of a gold-coated mirror to form a reference image at each investigated humidity level. The resulting spectral reflectance $R(\omega )$ of the object is calculated as

The reflectivity of the object $R_s(\omega )$ is related to the material refractive index $n(\omega )$ as

Analysis of the spectrum to identify spectral features and therefore classify the associated material can be performed manually by matching the features to materials of interest. This is the approach taken herein, for a proof-of-concept demonstration of the system using three materials. For a larger number of materials or mixtures of materials, this can become unwieldy. Automated classification using statistical methods, including unmixing, can be employed in these situations. This is addressed by a number of authors [14,16,17], which include using colormaps to show the spatial distribution of materials.

2.1 Telecentric lens

For the scan lens, an object-space telecentric f-theta design was chosen for its ability to maintain beam focus at an object over a wide range of scan angles and minimise changes in magnification with variations in object depth. The use of a single-mirror gimbal achieves coincidence of the two rotational axes of the mirror. When located in the conjugate plane of the lens, it minimises image distortions, maintains a good depth range of measurement and obviates the need for complex lens shapes. Polytetrafluoroethylene (PTFE) was selected as the lens material due to its machinability, availability in large sizes, relatively low THz absorption and dispersion, and a small refractive index of $n$ = 1.43 (offering low Fresnel losses) across a wide bandwidth (0.2-3 THz) [29].

Using ray tracing simulations, the 1 m reflection system design was optimised for minimum field curvature and distortion, with a hyperbolic shape for the first surface of the rotationally-symmetric lens, resulting in a lateral distortion of <1%. The curvature of each surface of the lens is given by $z = cr^2/ \left ( 1 + \sqrt {1 - (1 + k)c^2r^2} \right )$, where $c=1/R$ is the surface curvature, $R$ is the radius of curvature, $r$ is the radial coordinate and $k$ is the conic constant. For the lens herein, $R_1 = 452$ mm, $k_1 = -3.7$, $R_2 = -458$ mm and $k_2 = 0$. The maximum field angle for the 16 mm collimated beam diameter is 7.6$^{\circ }$ for a 160 mm diameter lens with a 150 mm clear aperture and a 20 mm center thickness. The lens parameters produce a diffraction-limited spot for all beam angles. The diffractive spot size is estimated at $\sim$16 mm FWHM for 0.45 THz, the peak of the dynamic range.

The field curvature was investigated experimentally by scanning the beam over a silver-coated mirror and recording the time delay of the reflected THz pulse over a grid of 81 $\times$ 81 pixels (2 mm step size). The 2-dimensional (2D) map (Fig. 2(a)) shows that the travel time of the beam gradually increases with increasing distance from the center of the lens, taking into account the double path the beam travels. The field curvature data is largely circularly-symmetric and has a maximum average field curvature between the center and black circle of only 0.75 ps, corresponding to a 225 $\mu$m depth offset. The circle outlines the maximum field angle to be applied without vignetting the hyperspectral image. The field map serves as a correction matrix applied to the images in the following section, to more precisely determine the time delay and hence the object depth at each pixel. Figure 2(b)) shows time-domain signals obtained from two maximum field angles (blue-bottom and red-left) as well as from the center (black), all showing identical pulse widths. The pulse width plays an important role in the spectral analysis as it determines the spectral response and hence the frequency range. A broadening of the pulse at different parts of the lens could lead to a reduced THz bandwidth and thus a non-uniform spectral response across the measured object. Additionally, the pulse amplitudes are maintained at the edges of the field, demonstrating that the beam is perpendicular to the object and therefore the system has good telecentric performance; without this, there can be a reduction in signal and thus dynamic range [22], leading to poor spectral performance.

 

Fig. 2. a) Field curvature for the telecentric lens with a clear aperture of 150 mm shown as a 2D time delay map. The beam travelling through the center of the lens is focused at the focal plane (0 ps) and requires less time to return to the detector than when travelling through outer parts of the lens. The black circle illustrates the maximum field angle that forms the scan area without vignetting. The blue and red dots are the positions of measurements taken at the maximum angle, to be compared with the pulse width at the center of the lens. b) Time-domain signals at the three locations. (These are not scaled for reflection and hence show twice the time delay.) The pulses are of identical width, inferring a homogeneous spectral behaviour across the lens.

Download Full Size | PDF

The losses at the scan lens are the sum of $\sim$0.14 dB Fresnel reflections at each material/air interface and a measured material absorption of 0.27 dB/cm for the applied THz spectrum. The power is therefore reduced by $\sim$1.64 dB at the center of the lens with the maximum thickness of 20 mm, considering the lens is passed twice. Adding the losses of 0.31 dB at each of the two TPX lenses as well as at the beamsplitter, $\sim$3.37 dB in reflection and $\sim$2.68 dB in transmission, the total system loss amounts to $\sim$8.31 dB.

3. Hyperspectral imaging

An aluminium block was machined to form a target with a spoke pattern to evaluate the system’s imaging performance. For spectral analysis, the ERC simulants lactose, tartaric acid and 4-ABA were included as pellets, located approximately 250 mm in front of the pattern. Signal averaging in the time domain is not required for pure imaging but increases the signal-to-noise ratio (SNR), which subsequently allows access to higher frequencies in the hyperspectral image. For the measurements herein, 100-times averaging was used. We additionally investigated the impact on the spectral analysis by reducing water vapor inside the enclosure through nitrogen purging, to increase the SNR further.

The gimbal can move faster between pixel positions than the THz spectral signals are generated and received. Hence, the scan time for hyperspectral images is only limited by the spectral system, which in this case is acquired at a rate of 8 Hz for a 100 ps time window. A faster TDS system could be employed, such as an electrically-scanned two-laser system [23], albeit with concomitant increases in cost and complexity. Conversely, the gimbal has programmable movement, allowing variable spatial resolution over the field of view. This way, fast full-field imaging can be combined with slower averaged spectral measurements at a few selected locations on the sample.

3.1 Imaging modality

A photograph of the measured objects is shown in Fig. 3. In the background is the metal spoke-pattern target, comprised of six segments having a depth of 10 mm each. In the foreground are the three simulant materials mounted on plastic supports. The hyperspectral image was acquired over the area represented by the red rectangle.

 

Fig. 3. Photograph (left) of several targets with a maximum height range of 255 mm. At the rear is a spoke target made from aluminium. In the foreground are three simulants mounted on a plastic ring. The 3D plot (right) of the THz height map shows the object heights at each point in the field of view.

Download Full Size | PDF

Based on the time delay data, a 3D relief image of the objects was extracted from the hyperspectral image by taking the time $t_{p1}$ of the first peak of the time-domain signal amplitude $E(t)$ for each pixel in the image. The feature depth $d$ at each pixel is then

The relief image is shown in Fig. 3, where the object features can be clearly identified based on their shape and depth. The data also contains information about the reflectivities of the materials (not shown here), which can help distinguish one material from another, e.g. metal vs dielectric.

The spoke-pattern can be used to determine the spatial resolution of the system. This was achieved by inscribing circles of increasing radius $r$ about the center of the pattern and taking the amplitude of their $N$^th harmonic, where $N$ is the number of spokes [22]. The corresponding arc length of one spoke is then $l = \pi r/N$, which is a measure of the minimum detectable feature size. In this case, $N = 6$ and $r=12$ mm, giving $l=6.3$ mm.

The time-step size of the TDS system is fixed at 33 fs or 10 $\mu$m in air, corresponding to a theoretical 5 $\mu$m depth ($\Delta$z) resolution. Such a fine depth resolution can be maintained over the field of view, for relative and absolute measurements, by incorporating the system correction matrix derived in Section 2.1. It is noted that the repeatability of the delay line is <5 fs peak-peak, which is negligible compared with the time increment.

The depth resolution was experimentally determined using a CNC machined aluminium target (Fig. 4) having seven regions with depths of 9, 20, 57, 97, 201, 508 and 1010 $\mu$m, as measured with a micrometer. The plots show the ability to resolve depth changes in the order of 10 $\mu$m. For the 508 $\mu$m step, the value measured from the depth map showed a 1.4% difference.

 

Fig. 4. Demonstration of micrometer-scale depth measurement using the surface relief data. a) A target with several depths was scanned and displayed as 2-dimensional map. b) Line scans across the relief map, at locations shown by the horizontal lines in a).

Download Full Size | PDF

3.2 Spectral modality

To demonstrate the spectral capabilities of the system, we extracted the time-domain information from the hyperspectral image of the three simulant materials in Fig. 3, taking the center pixel of each simulant pellet. A 100 ps window around the THz signal was then used for the spectral analysis. The resulting 10 GHz spectral resolution is suited to resolving the spectral features of the materials [14].

The regions-of-interest were identified and manually selected based on their spatial and depth features. Because the THz beam size is larger than the 2 mm step size of the image, a single pixel of the image is averaged over an area of the material. This spatial averaging helps to reduce noise and variances due to non-uniformities in the object. Near the boundaries of features within the object, the resulting signal is a convolution between the beam shape and the object features. This results in mixing of material spectra: for largely-featureless spectra such as polymers, the spectral features of the other material (material of interest) are preserved; for two materials of interest, spectral features are combined. For a discussion of spectral analysis and unmixing in these situations, see [16,17] and [14], which also incorporates water vapor effects.

Figure 5 summarises the results for the lactose pellet in both purged and unpurged environments. The top row shows the unpurged time-domain measurements (a) and corresponding frequency domain spectra (b) for lactose (red) and a prerecorded gold mirror reference (black), both at 55% relative humidity (RH). The resulting reflectance spectrum is shown in (c). The time-domain traces exhibit strong oscillations following the main peak due to sharp water vapor absorption lines, which are observed in the frequency domain spectra (b). The calculated reflectance spectrum, however, still allows identification of the characteristic spectral features of lactose at 0.54 and 1.38 THz [17]. The feature at 1.1 THz is due to water vapor [30].

 

Fig. 5. Time domain, frequency domain and reflectance data for lactose are shown in the top row measured at 55% RH, and the bottom row at 10% RH. Vertical dashed lines (c, f) identify characteristic spectral features. Despite strong water-vapor absorption, the lactose features at 0.54 and 1.38 THz can be observed at both humidity levels.

Download Full Size | PDF

The bottom row of Fig. 5 shows the nitrogen-purged results at 10% RH. The reduction of the water vapor increased the time-domain signal peak in (d) by 11% and reduced the noise after the peak, also reducing the water-vapor absorption lines in the frequency domain (e) and making the lactose spectral features clearer (f). The noise level increases significantly above $\sim$1.5 THz for non-purged measurements, while purged measurements extend the frequency range beyond this. However, the reflectance spectra demonstrate that two spectral features of lactose can be identified even without purging.

The reflectance spectrum of the tartaric acid sample is shown in Fig. 6(a)), where the prominent spectral feature can be found at the expected frequency of 1.1 THz, followed by the expected lower reflectivity thereafter [26]. Similarly, 4-ABA was identified through its characteristic spectral features at 0.60, 0.81 and 1.55 THz [10], as shown in Fig. 6(b)). Another reported spectral feature, at 1.30 THz [10], is not evident: the close-range transmission measurement therein shows the feature has low amplitude and close proximity to the strong feature at 1.55 THz; at the large standoff distance used herein, the feature becomes lost (as has been shown for other materials [26]). Despite this, three of the four frequencies are available for the identification of the substance.

 

Fig. 6. Reflectance spectra of a) tartaric acid, revealing the characteristic spectral feature at 1.1 THz and b) 4-ABA, showing characteristic spectral features at 0.60, 0.81 and 1.55 THz, both measured at 10% RH.

Download Full Size | PDF

4. Conclusion

We have demonstrated that THz hyperspectral 3D imaging can be performed with a compact fiber-coupled reflection system, representing a step towards practical field-deployable systems for security and industrial applications. The system produces broadband THz pulses, deriving spectral information through synchronous detection. Fast modulation of the signal in the time domain allows spectral acquisition at rates of 10 Hz or faster, while achieving spectral resolutions of 10 GHz. The beam is scanned laterally over an object of up to 150 mm, to build an image that contains information about both the reflectivity and height of the object’s features. The image thus contains temporo-spatio-spectral information, or a 3D hyperspectral image. By using a single-mirror scanner, images can be acquired quickly, without the more complex optical designs or extra signal processing required for two-mirror galvoscanner designs. The system can identify substances having spectral features up to 1.8 THz at a 1 m distance. Lactose, tartaric acid and 4-ABA could be identified in a normal room humidity environment, while optional purging with dry nitrogen gas improved the effective bandwidth of the measurements. The hyperspectral imaging system has proven to be insensitive to changes in object distance up to a range of 255 mm, which is currently the physical range of the delay line.

It is noted that absorption and reflection losses from the system components (quantified in Section 2.1) contribute to a reduction in dynamic range and hence bandwidth [22]. For the system herein, this limits spectral measurements to 1.8 THz, which is adequate to identify many materials. For larger bandwidth, the refractive optics could be replaced with reflective optics, and longer signal integration times could be used. The other major loss is due to the beamsplitter; removing this would necessitate non-normal incidence [26], affecting the depth-invariance of the system and complicating the beam scanning.

In some situations, larger standoff distances may be desirable. This can be achieved using a scan lens with a longer focal length. This will increase the spot size, reducing the spatial resolution, unless the diameter of the incident beam is increased, which in turn requires a larger lens, increasing the signal losses.

Another issue to consider is that of covering materials. While these might be separable from internal materials based on the time-domain data, where a narrow peak is used [27], they may affect spectral analysis due to overlap between the relatively wide time windows required to derive spectra with adequate spectral resolution [31]. Another consideration is scatter, either from covering materials or bulk materials [21,32]. This reduces the signal returned to the detector, requiring a higher dynamic range to compensate. Progress in the above areas will help to move the technology further towards practical in-field applications.