LINE‐SCAN HYPERSPECTRAL IMAGING PLATFORM FOR AGRO‐FOOD SAFETY AND QUALITY EVALUATION: SYSTEM ENHANCEMENT AND CHARACTERIZATION
M. S. Kim, K. Chao, D. E. Chan, W. Jun, A. M. Lefcourt, S. R. Delwiche, S. Kang, K. Lee
ABSTRACT. Line‐scan‐based hyperspectral imaging techniques have often served as a research tool to develop rapid multispectral methods based on only a few spectral bands for rapid online applications. With continuing technological advances and greater accessibility to and availability of optoelectronic imaging sensors and spectral imaging spectrographs, the range of implementation for hyperspectral imaging has been broadening across quality and safety inspection needs in the food and agricultural industries. We have developed a series of food inspection imaging systems based on hyperspectral line‐scan imaging with the use of a low‐light sensitive, electron‐multiplying charge‐coupled device (EMCCD). In this methodology article, the spectral and spatial system performance of the latest generation of the ARS hyperspectral imaging system, which is capable of reflectance and fluorescence measurements in the visible and near‐infrared (NIR) spectral regions from 400 to 1000 nm, is evaluated. Results show that the spectral resolution of the system is 4.4 nm at full‐width at half‐maximum (FWHM) and 6 nm FWHM at our typical operation mode (6-pixel spectral binning). We enhanced the system throughput responses by using spectral weighting filters to better utilize the dynamic range of the analog‐to‐digital converter. With this system throughput adjustment, noise‐equivalent reflectance measurements were significantly reduced by approximately 50% in the NIR region for a range of standard diffuse reflectance targets. The responsivity of the system from 450 to 950 nm was determined to be linear. Keywords. Fluorescence, Hyperspectral, Line‐scan imaging, Reflectance, Spectral calibration.
hemical and biological food properties can often techniques have been developed to combine the advantages be nondestructively assessed by spectroscopic of spectroscopy and machine vision in addressing agro‐food methods. Machine vision is already ubiquitous for quality and safety problems (Kim et al., 2001; Lu, 2003; Go‐ sorting items by their appearance, but convention‐ wen et al., 2007; Park et al., 2007). Hyperspectral imaging alC monochromatic or RGB‐based imaging methods em- methods provide full‐spectrum data, often hundreds of spec‐ ployed in conventional machine vision techniques are tral data points, for every pixel in the image of a food product, limited to evaluating basic physical attributes, such as the enabling spectral and spatial analysis for correlation to com‐ size, shape, and color of agricultural commodities (Aleixos position, contaminants, and physical attributes such as size et al., 2002; Blasco et al., 2007; Miller and Drouillard, 2001). and shape. However, high speeds and product volumes pres‐ In contrast, spectroscopic (i.e., hyperspectral) imaging can ent significant challenges to improving real‐time online allow for a thorough characterization of physical, chemical, spectral imaging inspection across agro‐food industries. and biological perturbations indicative of agro‐food product Fundamentally, there are two ways to acquire hyperspec‐ safety and quality. For these reasons, hyperspectral imaging tral imaging data from an object: the band sequential imaging method and the pushbroom (line‐scanning) imaging method. The band sequential imaging method captures a full spatial Submitted for review in July 2010 as manuscript number IET 8659; scene at each wavelength in a series of wavelengths to form approved for publication by the Information & Electrical Technologies a three‐dimensional hyperspectral image cube. The push‐ Division of ASABE in January 2011. broom (line‐scan) method captures a single line of spatial in‐ Company and product names are used for clarity and do not imply any formation containing full‐spectrum data for every spatial endorsement by USDA to the exclusion of other comparable products. pixel in the line, and the composite of a set of many spatial The authors are Moon S. Kim, Research Scientist, Kuanglin Chao, Research Scientist, Diane E. Chan, Agricultural Engineer, Won Jun, line‐scans forms a hyperspectral image cube. Research Associate, and Alan M. Lefcourt, ASABE Member Engineer, Since the early 2000s, applications of the hyperspectral Research Scientist, USDA‐ARS Environmental Microbial and Food Safety techniques as nondestructive means to assess safety and qual‐ Laboratory (EMFSL), Beltsville Agricultural Research Center, Beltsville, ity aspects of agricultural products have been increasing (Go‐ Maryland; Stephen R. Delwiche, Research Scientist, USDA‐ARS Food Quality Laboratory, Beltsville Agricultural Research Center, Beltsville, wen et al., 2007). Researchers at the USDA Agricultural Maryland; and Sukwon Kang, ASABE Member, Agricultural Engineer, Research Service (ARS) have developed several versions of and Kangjin Lee, Agricultural Engineer, National Institute of Agricultural hyperspectral imaging systems along with image analysis Engineering, Rural Development Administration, Suwon, Korea. techniques to address food safety and quality concerns for Corresponding author: Moon S. Kim, USDA‐ARS EMFSL, Beltsville food production and in food processing (Kim et al., 2001; Agricultural Research Center, Bldg 303 BARC‐East, 10300 Baltimore Ave., Beltsville, MD 20705‐2350; phone 301‐504‐8462; fax: 301‐504‐ Yang et al., 2006; Kim et al., 2008; Jun et al., 2009). The ARS 9466; e‐mail: [email protected]. hyperspectral imaging systems utilize the pushbroom line‐
Transactions of the ASABE Vol. 54(2): 703-711 2011 American Society of Agricultural and Biological Engineers ISSN 2151-0032 703 scan approach. Due to speed restrictions for data acquisition MATERIALS AND METHODS and processing, hyperspectral imaging has often been used as LINE‐SCAN HYPERSPECTRAL IMAGING SYSTEM a research tool to develop more rapid multispectral methods, The line‐scan hyperspectral imaging platform shown in based on only a few spectral bands, that can operate at higher figure 1 was designed for acquisition of reflectance measure‐ speeds for online applications (Kim et al., 2001). However, ments in the range of approximately 400 to 1000 nm and newer line‐scan‐based imaging technologies can effectively fluorescence measurements from approximately 420 to deliver high‐speed online safety and quality inspection of 700Ănm (with UV‐A excitation). Another consideration was food and agricultural products on high‐throughput process‐ flexibility for imaging objects of varying sizes. The platform ing lines in both hyperspectral and multispectral domains is able to accommodate imaging of individual grains of wheat (Chao et al., 2007). We have also developed a new line‐scan to whole chicken carcasses (i.e., adjustable field of view from hyperspectral imaging platform for high‐speed inspection on 80 to 300 mm). commercial processing lines that is capable of simultaneous multiple inspection algorithms for addressing different safety Sensing Components and quality problems (Kim et al., 2008). The line‐scan hyperspectral imaging system utilizes a In spite of the ubiquitous applications of hyperspectral low‐light sensitive electron‐multiplying charge‐coupled‐ imaging techniques and their development by independent device (EMCCD) camera (MegaLuca, Andor Technology, laboratories and/or commercial entities in agricultural fields, Inc., Belfast, Northern Ireland). The EMCCD has 1002 × details encompassing wavelength calibrations and character‐ 1004 pixels and is thermoelectrically cooled to ‐20°C via a ization of spectral‐dependent response are limited. Unlike two‐stage Peltier device. The EMCCD is equipped with a conventional spectroscopic systems, hyperspectral imaging 12.5 MHz pixel readout rate and 14‐bit A/D digitizer. Image captures spectral information at each spatial pixel location, data are transferred to a PC via USB 2.0. Both vertical and and the use of hypercube image data should not casually pro‐ horizontal pixels can be binned, and binning is achieved in ceed under the assumption that individual pixel responses are the hardware prior to data transfer to PC. equal. An imaging spectrograph (400 to 1000 nm, VNIR Hyper‐ In this methodology article, we present detailed system in‐ spec, Headwall Photonics, Inc., Fitchburg, Mass.) and a formation along with system characterization and illustrate C‐mount object lens (Schneider Optics, Van Nuys, Cal.) are a method to adjust spectral throughput to further enhance the attached to the EMCCD. The instantaneous field of view spectral‐spatial responses of the system. Characterization of (IFOV) is limited to a thin line by the spectrograph slit size. the system included wavelength calibration and examination A range of aperture slits (10, 25, 60, and 100 mm) is available, of the effects of pixel binning (in the spectral dimension) on and the slit can be readily changed. Typically, the 60 mm slit spectral bandwidth (resolution). Wavelength‐dependent re‐ is used most often in our system. Through the slit, light from sponse characteristics included (1) noise, in terms of signal‐ the scanned line of the IFOV is dispersed by the dispersive to‐noise ratio (SNR) or noise‐equivalent reflectance (NER), grating and projected onto the EMCCD. Therefore, for each and (2) responsivity, in terms of linearity with respect to line‐scan, a two‐dimensional (spatial and spectral) image is changes in exposure time for a fixed target, and linearity with created, with the spatial dimension along the horizontal axis respect to targets of a known range of reflectance intensities and the spectral dimension along the vertical axis of the using a fixed exposure time. EMCCD. Illumination Sources Two types of continuous‐wave (CW) light sources are used to provide the line‐scan imaging system's illumination for fluorescence and reflectance imaging. For reflectance, the source is a regulated 150 W DC quartz‐tungsten halogen
Figure 1. Illustrations of the laboratory‐based hyperspectral imaging system and critical sensing components.
704 TRANSACTIONS OF THE ASABE (QTH) lamp coupled with a pair of (bifurcated) fiber optic struments, Stratford, Conn.). To illuminate the entire line of line lights (Fiber‐Lite, Dolan‐Jenner Industries, Inc., Law‐ the IFOV, the Hg‐Ne pencil light was placed approximately rence, Mass.). To provide near‐nadir illumination to mini‐ 25 cm above (and at 5° forward angle) a square 30×30 cm mize shadowing, the two line lights are mounted to polytetrafluoroethylene (Spectralon) 99% diffuse white re‐ illuminate the line of the IFOV at backward and forward flectance reference panel (SRT‐99‐120, Labsphere, North angles of approximately 5°. The imaging system and its com‐ Sutton, N.H.) positioned horizontally level at the sample pre‐ ponents are encased in a light‐tight casing to mitigate the ef‐ sentation distance. Note that Ne is the starter gas for the spec‐ fect of ambient light. tral Hg‐Ne lamp; immediately upon supplying the power to For fluorescence imaging, two high UV‐A (320 to 400 nm, the lamp, the dominant Ne lines can be obtained first, fol‐ 365 nm peak) fluorescent lamp assemblies (Spectroline lowed by the Hg lines. XX‐15, Spectronics Corp., Westbury, N.Y.) are used. Be‐ System Characterization cause the presence of the QTH fiber‐optic line lights in the To test the system noise in the spectral regions of interest, confined space limited the options for placement of the UV six “known” laboratory standards, i.e., round 5 cm diameter fluorescent lamps, the UV lamps were mounted to illuminate Spectralon diffuse gray reflectance targets (Labsphere, North the line of the IFOV at angles of approximately 35° from the Sutton, N.H.) for 5%, 10%, 20%, 40%, 60%, and 80% reflec‐ vertical position, backward and forward. A long (>400 nm) tance, were used with the 30×30 cm 99% diffuse white re‐ wavelength pass filter (Kodak Wratten 2A gelatin filter) was flectance panel for reference. Ten hyperspectral image cubes placed between the C‐mount lens and the spectrograph to re‐ of the set of six standard targets were acquired. For deter‐ move the second‐order effects in the 700 to 760 nm (caused mination of SNR, 20 independent scans using the 99% reflec‐ by the overlapping of the grating second‐order diffracted tance target in a fixed position were acquired. For each scan, beam) of the UV fluorescence excitation sources. a respective dark current measurement was acquired and sub‐ System Operations sequently subtracted from the sample measurement. Interface software was developed, using a software devel‐ The wavelength‐dependent throughput of the hyperspec‐ opment kit provided by the camera manufacturer, on a Micro‐ tral imaging system can be affected by light source irra‐ soft Visual Basic (version 6.0) platform in the Microsoft diance, the optical components, the spectrograph, and the Windows operating system. Because of the pushbroom quantum efficiency (QE) of the CCD sensor. In order to based, line‐scan imaging method coupled with the use of the quasi‐optimize the wavelength‐dependent system through‐ dispersive spectrograph, the spectral information at each pix‐ put and to use a greater portion of the 14‐bit dynamic range el location is presented in the vertical (traverse) direction of of the A/D digitizer, the system throughput was weighted by the EMCCD. With the aid of the software controls for the filters (Kodak Wratten 82C gelatin filters). A comparison of EMCCD, the system can be configured to acquire either hy‐ the system throughput enhancement in terms of NER of the perspectral images (with the entire span of the vertical spec‐ standard targets from 400 to 1000 nm was performed. tral region) or multispectral images (with a few selected regions of interest in the vertical spectral direction using the random track acquisition mode). Vertical binning affects the RESULTS AND DISCUSSION spectral interval and, more importantly, the spectral resolu‐ tion of the images. In general, reducing the data volume per SPECTRAL CALIBRATION The image in figure 2 shows a line‐scan image of the Hg line‐scan image results in a higher number of lines scanned emission lines acquired with no pixel binning (1004 × 1002 per second. However, the pixel readout and A/D conversion m rates are limiting factors, along with the EMCCD exposure pixels) and with a 60 m slit. The adjacent graph on the right shows the prominent Hg emission lines at 404.3, 435.8, time, for the number of lines scanned per second. 546.7, and 578.0 nm and the corresponding vertical (spectral) During image acquisition, a sample is moved transversely through the illuminated IFOV by a programmable motorized y‐pixel locations at the 501st horizontal (spatial) x‐pixel loca‐ tion indicated in the image. The two emission lines observed translation stage (Velmex, Inc., Bloomfield, N.Y.) that is con‐ between 600th and 800th y‐pixel locations show the second‐ trolled through the interface software. With each new line‐ scan acquisition, the translation stage is advanced by one order effect of the 404.3 and 435.8 nm Hg emission lines. An optimal alignment of the EMCCD‐spectrograph results in the incremental step. Depending on the nature of the samples be‐ same vertical pixel locations for the individual Hg emission ing imaged, the step sizes may be as small as <0.1 mm for small target samples when higher spatial resolution images lines regardless of the horizontal pixel location, as seen in fig‐ ure 2. If the EMCCD and spectrograph were to be misaligned, are needed, or as large as several millimeters when fine spa‐ skewed emission lines (off the horizontal) would be ob‐ tial resolution is not necessary. In addition to providing precise control of the incremental served. Note that crescent‐shaped emission lines may be ob‐ served in the presence of an artifact known as the “smiley movement of a sample for line‐scan imaging, the translation effect.” If observed, this may require spectral calibration for stage also enables the system to perform repeated imaging of a target sample with no change in the spatial position of the individual spatial pixel locations (Lawrence et al., 2003). Figure 3a shows representative Hg and Ne emission lines sample within different images, allowing for direct spatial and the wavelength calibration result. With the known Hg comparison between images as needed, e.g., to examine the spectral characteristics of a particular sample pixel imaged and Ne emission line wavelengths as the dependent variable and the corresponding vertical pixel locations as the indepen‐ using variations in exposure time or illumination. dent variables, a linear or polynomial regression fit is typical‐ Spectral Calibration ly used to determine the wavelengths corresponding to the Spectral calibration of the EMCCD‐spectrograph was per‐ vertical pixel locations. The linear fit resulted in an r2 value formed using an Hg‐Ne spectral calibration lamp (Oriel In‐ of 0.999 with the vertical pixels spaced by approximately
Vol. 54(2): 703-711 705 Figure 2. Line‐scan image of Hg lamp emission lines acquired with no pixel binning (1004 × 1002 pixels). The horizontal (x) axis represents the spatial dimension, and the vertical axis (y) represents the spectral dimension. The Hg emission spectrum across the 501st horizontal pixel location is shown on the adjacent graph.
Figure 3. (a) Representative Hg and Ne emission spectra and linear regression fit for wavelength calibration, and (b) residual errors for the linear re‐ gression fit.
0.79 nm. Figure 3b shows the residual errors of the linear re‐ Note that the data used to determine the spectral resolution gression fit. The result here exhibits a random pattern in the were acquired with the use of a 60 mm spectrograph slit, and residual errors (less 1 nm), suggesting that a linear regression the slit size in conjunction with the vertical binning affects fit for wavelength calibration was sufficient. the spectral resolution. We typically operate the imaging sys‐ tem with 6× vertical pixel binning, which results in spectral SPECTRAL RESOLUTION data of approximately 4.8 nm intervals with a spectral resolu‐ The effects of spectral pixel binning on spectral resolution tion of approximately 6.0 nm. This adequately represents the have been seldom discussed in the literature. Without any generally broad spectral characteristics of our agricultural vertical spectral pixel binning, a hyperspectral image cube in samples. Note also that the absorption (depicted in reflec‐ the 400 to 1000 nm range consists of over 800 channels. This tance) or fluorescence emission characteristics of plant‐ and may result in redundancy of spectral information and the cre‐ animal‐based samples in the 400 to 1000 nm range are rela‐ ation of unnecessarily large data files, since biological mate‐ tively broad (i.e., FWHM) in nature. Binning reduces redun‐ rials do not exhibit such optical properties or fluorescence dant spectral data and unnecessarily large data volumes. The emissions as to warrant such large‐dimension spectral data. 6× binning yields approximately 5 nm interval data, which Figure 4 shows a comparison of the Hg lines acquired with allows for adequate spectral representation of agricultural no vertical binning and with 6× vertical binning. Using the samples, which rarely exhibit sharp features such as those in Hg emission line at 546.7 nm, we determined the spectral res‐ figure 4. For fluorescence imaging, the wavelengths span olution in full‐width at half‐maximum (FWHM) for no verti‐ from approximately 420 to 700 nm across 60 wavebands, cal pixel binning (1×) and then for 3×, 6×, and 10× vertical while reflectance images are captured from approximately pixel binning (fig. 4 inset). The spectral responses of individ‐ 400 to 1000 nm with 125 wavebands. ual binning selections were subjected to a cubic‐spline inter‐ polation to obtain spectral responses in 0.1 nm increments, SPATIAL RESOLUTION and then the wavelength differences (FWHM) at 50% of the A myriad of factors can affect the spatial resolution of the peak intensity levels before and after the peak wavelength line‐scan imaging system. Spatial resolution in the direction were calculated. The FWHM values were 4.4, 4.6, 6.0, and of the translation stage movement is dictated by the line‐scan 9.4 nm for 1×, 3×, 6×, and 10× binning, respectively. step increment. As shown in figure 5a, effective line‐scan
706 TRANSACTIONS OF THE ASABE Figure 4. Effect of pixel binning on the spectral dimension showing the comparison of the Hg lines acquired with no vertical binning and 6× vertical binning. The inset graph shows full‐width at half‐maximum (FWHM) for no vertical pixel binning (1×) and for 3×, 6×, and 10× vertical pixel binning for the Hg emission line at 547.7 nm. All spectral responses were subjected to cubic‐spline interpolation.
Figure 5. (a) Three scenarios that can affect the spatial resolution of line‐scan image acquisition, and (b) U.S. Air Force resolution and distortion stan‐ dard target images at 650 nm, acquired with 0.08 mm step increment of the translation stage to provide a near‐square pixel size of 0.08 × 0.08 mm. The third target image is an enlargement of the middle portion of the distortion standard target. imaging would cover the entire sample surface with allow‐ coarse‐resolution images and can misrepresent samples that ance for minor overlap between successive line‐scan images exhibit relatively small spatial features. Spatial resolution in (minor oversampling). Undersampling, by scanning with the transverse direction (along length of IFOV) is a combined step increments greater than the IFOV width, will result in function of the detector pixel size and the IFOV length.
Vol. 54(2): 703-711 707 Changing the distance between the object lens and the target blue‐green and NIR regions compared to the red region to affects both the IFOV length and the IFOV width. The width better balance the responses. Note that the use of 82C filters is also affected by the spectrograph slit size. required a longer exposure time, typically twice as long. After adjusting the lens‐to‐sample distance to capture a However, because of the use of the low‐light sensitive full‐scene IFOV length of approximately 80 mm, line‐scan EMCCD, an exposure time of less than 10 ms was required images of standard targets (USAF 1951 resolution standard for a typical reflectance scan. and distortion target, Edmund Optics, Barrington, N.J.) at 650 nm, shown in figure 5b, were acquired with 0.08 mm step CHARACTERIZATION OF SYSTEM NOISE increments of the translation stage to provide a near‐square In order to characterize the system performance, we eval‐ pixel size of 0.08 × 0.08 mm (and 1004 × 1000 pixels). Note uated a noise‐equivalent change in reflectance (NER) with that the width of the IFOV was determined to be approxi‐ and without the 82C filters at a fixed exposure time. NER was mately 0.11 mm, thereby causing slight oversampling of the calculated by dividing the standard deviation by the mea‐ target. The image on the right in figure 5b is an enlargement sured reflectance values of the standard 5%, 10%, 20%, 40%, of the middle portion of the image of the standard target, 60%, and 80% reflectance targets. In addition, the signal‐to‐ showing that the individual dots, 0.25 mm in diameter and noise ratio (SNR) of the system for a range of peak intensity spaced 0.50 mm apart, can be clearly resolved with minimal response levels, from 10% to 90% of the peak (saturation) distortion. level in 10% increments, was evaluated with the use of 82C filters by proportionally adjusting exposure times. SNR was THROUGHPUT ADJUSTMENTS calculated as the ratio of ten individual scan averages of the Ideally, flat and uniform responses in the spectral and spa‐ system to the standard deviation of those ten measurements. tial domains, respectively, are desirable in a hyperspectral Figure 7a shows a reflectance image of the six standard imaging system. Unfortunately, a myriad of artifacts emanat‐ targets at 650 nm acquired with 82C filters. Note that for the ing from the combination of the system components along the ten repetitions of line‐scan hyperspectral measurement of the light path, such as light irradiance, optics, and EMCCD QE, targets with and without 82C filters, the arrangement of the affect the overall system responses. Wavelength regions out‐ standard targets was never changed. Figure 7b shows a com‐ side of the peak response range are limited in the use of the posite image of ten line‐scans of the 99% reference target in full A/D dynamic range of the digitizer and can have the ef‐ a fixed position using nine progressive decreases in exposure fect of potentially reducing SNR. Traditional photography time from that of the saturation level at 650 nm. utilizes color filters to perform photometric correction. We The resultant reflectance spectra shown in figure 8 were examined a number of readily available color filters and extracted from a 10×10 pixel region of interest (ROI) within found the Kodak Wratten 82C gelatin filter (transmittance the image areas of the standard reflectance targets, acquired shown in fig. 6a) suitable for providing wavelength‐weighted with and without 82C filters. In the peak response wavelength correction suitable to adjust the throughput of our system. regions of the system, minimal effects of the 82C filters were With the use of the 82C filters, approximately 40% and observed. The most noticeable differences were seen in the 70% of the light in the blue‐green and NIR regions, respec‐ blue (400 to 500 nm) and NIR (from 800 to 1000 nm) regions tively, was transmitted while limiting the transmission of the of the spectrum, where the reflectance was underestimated peak throughput region to approximately 20%. Figure 6b il‐ without the use of 82C filters. Reduction of a relatively great‐ lustrates representative system responses from 400 to er fraction of the throughput in the blue region compared to 1000Ănm without the 82C filter and with one and two 82C fil‐ the NIR region by the 82C filters (e.g., reduced throughput ters. The filters were inserted between the camera lens and including approximately 40% of original blue transmission the spectrograph. The gelatin filters are thin and flexible, and vs. 70% of original NIR transmission) might have also re- they can be easily cut into a circular shape to fit. We opted to duced the spectrograph‐produced second‐order effect ob‐ use a stack of two 82C filters to increase the throughput in the served in the NIR.
Figure 6. (a) Transmittance characteristics with use of one and two 82C filters, and (b) representative system responses from 400 to 1000 nm without the 82C filter and with one and two 82C filters. Responses were obtained by imaging the 99% diffuse reflectance target in reflectance mode.
708 TRANSACTIONS OF THE ASABE Figure 7. Reflectance image of the six standard reflectance targets at 650 nm, acquired with 82C filters, and (b) composite image of ten line‐scans of the 99% reflectance reference target acquired with nine progressive decreases in exposure time.
flectance targets. Overall, the NER values are relatively small for the standard targets, suggesting that regardless of the throughput adjustment, the system is viable for hyper‐ spectral imaging in the spectral regions under investigation. However, significant reduction of NER values in the NIR re‐ gion, from about 750 to 1000 nm, was noted for the reflec‐ tance targets. Individual spatial pixel‐based SNR values at each wave‐ length were determined for the set of ten exposure times (fig.Ă7b). The figure 10 image (right) shows the spatial pixel‐ by‐pixel SNR values at the 50% peak response exposure level from approximately 400 to 1000 nm. The adjacent plots (left) depict wavelength‐dependent SNR values at three spatial pixel locations: 251, 551, and 751. The mean (pixel 1 to 1004) SNR values at each of the exposure levels resembles that of the raw digital responses. Even at 10% of the saturation level, the mean SNR of the system exceeded 100, except for wave‐ lengths shorter than approximately 450 nm. The use of the second‐order blocking filter (Kodak Wratten 2A) for fluores‐ cence measurements reduced the throughput of the system in this region, and the intensity response values approached that of the dark current level toward 400 nm. For sample evalua‐ tions, we typically acquire sample reflectance scans from approximately 450 nm and above. Linear responsivity of the system from 450 to 950 nm in 100 nm increments was also determined using the six stan‐ Figure 8. Mean reflectance spectra from a 10×10 pixel region of interest dard reflectance targets with a fixed exposure time and the of the standard diffuse reflectance targets, without the 82C filter and with 99% reflectance reference panel with the range of exposure two 82C filters. times. The results (not shown for brevity) indicated that the hyperspectral system responded linearly as a function of Figure 9 illustrates the resultant effect of the 82C filters on changes in target brightness and exposure time, with a mini‐ NER for the 5%, 10%, 20%, 40%, 60%, and 80% standard re‐ mal r2 value of 0.999 from 450 to 950 nm.
Vol. 54(2): 703-711 709 0.020 0.020 0.12 80 % Standard 0.018 0.018 60 % Standard 40 % Standard
0.016 0.016 0.10 W Wratten 82C filters W Wratten 82C filters W Wratten 82C filters 0.014 W/O 82C filter 0.014 W/O 82C filter W/O 82C filter 0.08 0.012 0.012
0.010 0.010 0.06 % Reflectance) % Reflectance) % Reflectance) D D 0.008 0.008 D 0.006 0.04 0.006 NER ( NER ( NER ( 0.004 0.004 0.02 0.002 0.002 0.000 0.000 0.00 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000
Wavelength (nm) Wavelength (nm) Wavelength (nm)
0.16 0.6 3.0 20 % Standard 0.14 10 % Standard 5 % Standard 0.5 2.5 0.12 W Wratten 82C filters W Wratten 82C filters W Wratten 82C filters W/O 82C filter 0.4 W/O 82C filter 0.10 2.0 W/O 82C filter
0.08 0.3 1.5 % Reflectance) % Reflectance) % Reflectance) D D 0.06 D 0.2 1.0 NER ( NER ( 0.04 ( NER
0.1 0.5 0.02
0.00 0.0 0.0 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000
Wavelength (nm) Wavelength (nm) Wavelength (nm)
Figure 9. Noise‐equivalent reflectance (n = 10) of the standard diffuse reflectance targets, without the 82C filter and with two 82C filters.
Figure 10. Signal‐to‐noise ratio image at 50% of the peak response levels. The image depicts SNR at individual spatial pixels from approximately 400 to 1000 nm. The adjacent plots (left) show the SNR at three spatial pixel locations indicated in the image (251, 551, and 751).
CONCLUSIONS spectral dimension of a two‐dimensional array detector in a A myriad of artifacts emanating from the combination of hyperspectral imaging system has a significant effect on the system components along the light path, such as illumination spectral resolution of the images acquired. It was determined source, optics, and EMCCD QE, can affect the performance that with 6‐pixel binning in the spectral dimension, as is typi‐ of imaging systems. We evaluated the spectral and spatial cally used for our system operation, the spectral resolution system performance of the latest generation of the ARS hy‐ was 6 nm at FWHM. perspectral imaging system in the visible and near‐infrared We devised the use of gelatin color filters as spectral spectral regions from 400 to 1000 nm. Pixel binning in the weighting filters to adjust the throughput and, consequently,
710 TRANSACTIONS OF THE ASABE to better utilize the dynamic range of the analog‐to‐digital Gowen, A. A., C. P. O'Donnell, P. J. Cullen, G. Downey, and J. M. converter. With this throughput adjustment, NER measure‐ Frias. 2007. Hyperspectral imaging: An emerging process ments were significantly reduced by approximately 50% in analytical tool for food quality and safety control. Trends in the NIR region for a range of standard diffuse reflectance tar‐ Food Sci. and Tech. 18(12): 590‐598. gets. The SNR of the system at individual pixels was deter‐ Jun, W., M. S. Kim, K. Lee, P. Millner, and K. Chao. 2009. Assessment of bacterial biofilm on stainless steel by mined to be over 100 from 450 to 1000 nm. The responsivity hyperspectral fluorescence imaging. Sensing and Instrum. for of the system in the spectral range from 450 to 950 nm was Food Quality and Safety 3(1): 41‐48. also determined to be linear. Kim, M. S., Y. R. Chen, and P. M. Mehl. 2001. Hyperspectral reflectance and fluorescence imaging system for food quality ACKNOWLEDGEMENTS and safety. Trans. ASAE 44(3): 721‐729. This work was partially supported by a grant from the Bio‐ Kim, M. S., K. Lee, K. Chao, A. M. Lefcourt, W. Jun, and D. E. Green 21 Program (No. PJ007208), Rural Development Ad‐ Chan. 2008. Multispectral line‐scan imaging system for ministration, Suwon, Korea. simultaneous fluorescence and reflectance measurements of apples: Multitask apple inspection system. Sensing and Instrum. for Food Quality and Safety 2(2): 123‐129. Lawrence, K. C., B. Park, W. R. Windham, and C. Mao. 2003. REFERENCES Calibration of a pushbroom hyperspectral imaging system for Aleixos, N., J. Blasco, F. Navarrón, and E. Moltó. 2002. agricultural inspection. Trans. ASAE 46(2): 513‐521. Multispectral inspection of citrus in real‐time using machine Lu, R. 2003. Detection of bruises on apples using near‐infrared vision and digital signal processors. Computers and Electronics hyperspectral imaging. Trans. ASAE 46(2): 523‐530. in Agric. 33(2): 121‐137. Miller, W. M., and G. P. Drouillard. 2001. Multiple feature analysis Blasco, J., N. Aleixos, J. Gómez‐Sanchis, and E. Moltó. 2007. for machine vision grading of Florida citrus. Applied Eng. in Citrus sorting by identification of the most common defects Agric. 17(5): 627‐633. using multispectral computer vision. J. Food Eng. 83(3): Park, B., W. R. Windham, K. C. Lawrence, and D. P. Smith. 2007. 384‐393. Contaminant classification of poultry hyperspectral imagery Chao, K., C. C. Yang, Y. R. Chen, M. S. Kim, and D. E. Chan. using a spectral angle mapper algorithm. Biosystems Eng. 96(3): 2007. Hyperspectral/multispectral line‐scan imaging system for 323‐333. automated poultry carcass inspection applications for food Yang, C. C., K. Chao, Y. R. Chen, M. S. Kim, and D. E. Chan. safety. Poultry Sci. 86(11): 2450‐2460. 2006. Development of fuzzy logic based differentiation algorithm and fast line‐scan imaging system for chicken inspection. Biosystems Eng. 95(4): 483‐496.
Vol. 54(2): 703-711 711 712 TRANSACTIONS OF THE ASABE