Optimal Optical Filters of Fluorescence Excitation ... - ScienceCentral

2 downloads 0 Views 2MB Size Report
Optimal Optical Filters of Fluorescence Excitation and Emission ... 2Department of Biosystems Machinery Engineering, College of Agricultural and Life Science, ...
Original Article

Journal of Biosystems Engineering

J. of Biosystems Eng. 37(4):265-270. (2012. 8) http://dx.doi.org/10.5307/JBE.2012.37.4.265

eISSN : 2234-1862 pISSN : 1738-1266

Optimal Optical Filters of Fluorescence Excitation and Emission for Poultry Fecal Detection 1

2

3

2

2*

Taemin Kim , Hoonsoo Lee , Moon S Kim , Wang-Hee Lee , Byoung-Kwan Cho 1

Intelligent Robotics Group, NASA Ames Research Center, MS 269-3, Moffett Field, CA 94035, USA Department of Biosystems Machinery Engineering, College of Agricultural and Life Science, Chungnam National 3 University, 220 Gung-Dong, Yusung-Gu, Daejeon, Korea, Environmental Microbial and Food Safety Laboratory, Animal and Natural Resources Institute, Agricultural Research Service, United States Department of Agriculture, 10300 Baltimore Ave, Beltsville, MD 20705, United States 2

Received: June 25th, 2012; Revised: August 3rd, 2012; Accepted: August 20th, 2012

Purpose: An analytic method to design excitation and emission filters of a multispectral fluorescence imaging system is proposed and was demonstrated in an application to poultry fecal inspection Methods: A mathematical model of a multispectral imaging system is proposed and its system parameters, such as excitation and emission filters, were optimally determined by linear discriminant analysis (LDA). An alternating scheme was proposed for numerical implementation. Fluorescence characteristics of organic materials and feces of poultry carcasses are analyzed by LDA to design the optimal excitation and emission filters for poultry fecal inspection. Results: The most appropriate excitation filter was UV-A (about 360 nm) and blue light source (about 460 nm) and band-pass filter was 660-670 nm. The classification accuracy and false positive are 98.4% and 2.5%, respectively. Conclusions: The proposed method is applicable to other agricultural products which are distinguishable by their spectral properties. Keywords: Excitation and emission filters, Multispectral imaging model, Poultry feces, Spectrofluorimetry

Introduction Automated inspection of live and slaughtered poultry has been requested as poultry production and consumption increased (Bilgili, 2001). Each poultry carcass slaughtered and processed at poultry plants in the U.S is inspected to comply with the Pathogen Reduction/Hazard Analysis and Critical Control Point (PR/HACCP) rule. Human inspectors to inspect individual chicken carcasses on the processing lines have a tendency to develop repetitive motion injuries and fatigue problems in noisy and highly humid environment (OSHA, 1999). Therefore, market forces are at work encouraging the use of more sophisticated food safety technology along with an *Corresponding author: Byoung-Kwan Cho Tel: +82-42-821-6715; Fax: +82-42-823-6246 E-mail: [email protected]

expanded array of food safety practices (Food Safety and Inspection Service, 2006). Computer vision, hyperspectral imaging, and optical systems for poultry inspection are prevailed to discriminate wholesome from unwholesome chicken carcasses (Chao et al., 2002). Hyperspectral imaging techniques have been utilized in many scientific disciplines, from microscopic studies to airborne remote-sensing applications (Kim et al., 2004). A hyperspectral data is a threedimensional data containing two-dimensional information measured at a sequence of individual wavelengths across a sufficiently broad spectral range. The resultant spectra can be used, in principle, to characterize and identify any given material. Recently, a laboratory-based hyperspectral imaging system which employs a pushbroom method was developed (Kim et al., 2004). Compliance with zero tolerance of feces in poultry

Copyright ⓒ 2012 by The Korean Society for Agricultural Machinery This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Kim et al. Optimal Optical Filters of Fluorescence Excitation and Emission for Poultry Fecal Detection Journal of Biosystems Engineering • Vol. 37, No. 4, 2012 • www.jbeng.org

processing establishments is currently verified by visual observation. Automated fecal inspection systems that can operate on-line in real-time in the slaughter plant environment are needed. A hyperspectral imaging system for detecting fecal and ingesta contaminates was demonstrated (Lawrence et al., 2001). The ratio of the 565-nm image divided by the 517-nm image was able to identify fecal and ingesta contaminates. Visible, near infrared reflectance spectroscopy was investigated to discriminate between uncontaminated poultry breast skin and feces, and to select key wavelengths for use in a hyperspectral system (Kim et al., 2003). Four preprocessing methods are compared to analyze hyperspectral data considering two parameters: calibration and 20-nm spectral smoothing and showed that a band ratio with 2-wavelength equation (565/517) performed very well (Park et al., 2002). The principal component analysis technique was employed to finds an effective representation of spectral signature in a reduced dimensional feature space and a support vector machine to makes a decision whether each pixel falls in normal or tumor categories (Fletcher and Kong, 2003). A method for detecting skin tumors on chicken carcasses was proposed using hyperspectral fluorescence imaging data (Kim et al., 2004). The objective of this study is to develop a new analysis method to design optimal excitation and emission filters in a multispectral fluorescence imaging system for poultry fecal inspection. A new method to design excitation and emission filters of a multispectral fluorescence imaging system for poultry fecal inspection is proposed. The optimal excitation and emission filters are designed for poultry fecal inspection through linear discriminant analysis (LDA). A mathematical model for hyperspectral imaging system is proposed and its system parameters, such as excitation and emission filters, were optimally determined by LDA. An alternating scheme was proposed for numerical implementation. Spectrofluorimetric data of organic materials and feces of chicken carcasses were analyzed by LDA. The optimal excitation and emission filters of a multispectral imaging system for poultry fecal inspection were consistent with what experts provided in previous research (Cho and Kim, 2007). The proposed method is applicable for other agricultural products which are distinguishable by their spectral properties. The paper is organized as follows: in material and methods a mathematical formulation of multispectral imaging models is provided, and its properties are

investigated; optimal light source and filter obtained by LDA is presented in result and discussion; and concluding remarks are made in conclusion.

Materials and Methods Sample Spectrofluorimetric responses of poultry feces (intestines, duodenum, small intestine, cecum, rectum) and organic materials (blood, skin, and flesh) were measured. The digestive tracts were eviscerated from chicken carcasses and their feces gathered including skin and flesh. Spectrofluorimetric responses of specimens put into a cuvette were measured by a spectrofluorimeter (Fluorolog III, Horiba Industries, Edison, N.J., USA). The light source was excited from 350 nm to 650 nm by 5 nm steps and the emissions were measured from 350 nm to 750 nm by 2 nm steps. MATLAB software was used to calculate discriminability from spectral data of eight specimens and to obtain the optimal excitation and emission filters

Multispectral imaging model The multispectral imaging system for spectrofluorimetry of poultry specimen consists of light sources, excitation and emission filters, and cameras (Fig. 1). A light source is assumed to be white and a camera has the uniform sensitivity for all wavelengths. Suppose that a specimen shows its own hyperspectral response h(u,v) with random noise n(u,v). Its spectrofluorimetric response is:

r (v) = rs (v) + rn (v),

where rs (v) := ∫ s(u )h(u, v)du and rn (v) = ∫ s(u )n(u, v)du

are a signal and a noise of spectrofluorimetric response, respectively. The noise characteristic of hyperspectral response was investigated. The noise is n(u,v) assumed to be a Gaussian random noise:

n(u , v) ~ N (0, σ 2 (u , v)),

(2)

2

where σ (u,v) is variance at (u,v). The sample mean and variance of spectrofluorimetric response are calculated by: r (v) = E[r (v)] = rs (v)

σ 2 (v) := Var[r (v)]

(3) ∞

= Var[rn (v)] = ∫ s 2 (u )σ 2 (u , v)du. 0

266

(1)

Kim et al. Optimal Optical Filters of Fluorescence Excitation and Emission for Poultry Fecal Detection Journal of Biosystems Engineering • Vol. 37, No. 4, 2012 • www.jbeng.org

(a)

(b)

Figure 1. Multispectral imaging model and fluorescent spectrum; Multispectral imaging model (a), A typical fluorescence emission and excitation matrix in relative fluorescence intensity (RFI).

The intensity g through a filter f(v) is:

c

g = ∑ pi gi

(7)

i =1

g = ∫ f (v)r (v)dv

(4)

= gs + gn ,





where g s := f (v)rs (v)dv and g n := f (v)rn (v)dv are a signal and noise of intensity, respectively. The sample mean and variance of intensity are obtained by:

g := E[ g ] = g s

where pi is prior probability of the ith specimen. The within and between variances are obtained by (Duda et al., 2001): c

SW ( f , s ) := ∑ piσ i2 i =1

⎡ ⎤ ⎧ c ⎫ = ∫ ∫ ⎢ s 2 (u ) ⎨∑ piσ i2 (u , v ) ⎬ f 2 (v ) ⎥ dudv, i 1 = ⎩ ⎭ ⎣ ⎦ c

= ∫ ∫ s(u )h(u, v) f (v)dudv

σ 2 := Var[ g ] = Var[ gn ] =∫



0





0

S B ( f , s ) := ∑ pi ( g i − g ) i =1

(5)

s (u )σ (u, v) f (v)dudv 2

2

2

For multiple specimens the sample mean and variance of the intensity of the ith specimen are

σ i2 := Var[ g | ωi ] =∫



0





0

c

= ∑ pi ⎡ ∫ ⎢⎣ 0 i =1

(6)

s 2 (u )σ i2 (u , v) f 2 (v)dudv,

2  and σi are mean and where ωi is the ith specimen, 

variance of spectrofluorimetric response of ωi. The total mean of intensity is:







0

2

s (u ){hi (u , v ) − h (u , v )} f (v ) dudv ⎤ . ⎥⎦

The excitation and emission filters should be chosen to maximize the discriminant power of specimens. The discriminability in LDA is defined by

J ( f , s) =

g i = E[ g | ωi ] = ∫ ∫ s (u )hi (u , v) f (v) dudv,

(8)

2

S B ( f , s) SW ( f , s)

(9)

The discriminability varies with form of f and s but not scalar product. Their function space is restricted to positive unit functions:

( f ∗ , s∗ ) = arg max J ( f , s) f , s∈B (R + )

(10)

+

where B(R ) is a collection of all positive unit functions. The optimal excitation and emission filter is obtained numerically.

267

Kim et al. Optimal Optical Filters of Fluorescence Excitation and Emission for Poultry Fecal Detection Journal of Biosystems Engineering • Vol. 37, No. 4, 2012 • www.jbeng.org

Results and Discussion The representative excitation and emission plot of poultry skin and rectal feces are shown in Fig. 2. Continuous f and s were discretized by the same resolution, initialized with constant functions, and alternatingly optimized. The optimal relative attenuation values of excitation and emission filters are shown in Fig. 3 (a) and (b), respectively. The proposed method provides continuous forms, while previous research (Cho and Kim, 2007) experimentally presented selective bandwidths. An excita-

tion filter composed of UV-A (about 360 nm) and blue light (about 460 nm) and a band-pass filter with 660-670 nm bandwidth were most appropriate. These results were consistent with qualitative studies (Cho and Kim, 2007) and provide a quantitative optimal excitation and emission filters. The intensity values of poultry feces and organic materials provide clear distinction so that single threshold value is sufficient to decision. The intensity values are obtained by applying excitation and emission filters to the spectrofluorimetric responses of poultry feces and organic

(a)

(b)

Figure 2. Fluorescent spectra of poultry; Fluorescent spectrum of poultry skin (a), Fluorescent spectrum of poultry feces (b).

(a)

(b)

Figure 3. Optimal excitation and emission filters; excitation Filter (a), emission Filter (b).

Figure 4. Class histogram of poultry feces and organic materials.

268

Kim et al. Optimal Optical Filters of Fluorescence Excitation and Emission for Poultry Fecal Detection Journal of Biosystems Engineering • Vol. 37, No. 4, 2012 • www.jbeng.org

Figure 5. Spectrofluorimetric responses of samples.

Table 1. Classification rate Decision

Poultry Feces

Organic Materials

Total

Poultry Feces

39 (98.4%)

1 (2.5%)

40 (100%)

Organic Materials

0 (0%)

24 (100%)

24 (100%)

Total

39

25

64

Truth

269

Kim et al. Optimal Optical Filters of Fluorescence Excitation and Emission for Poultry Fecal Detection Journal of Biosystems Engineering • Vol. 37, No. 4, 2012 • www.jbeng.org

materials. Their histograms overlap rarely so that one observation of poultry feces overlaps with organic materials (Fig. 4). The classification accuracy is 98.4% and false positive is 2.5% if the threshold value is set by 8 × 107. The false positive is duodenum sample number 8 as it does not have common feature of poultry feces (Fig. 5).

Conclusions The optimal excitation and emission filters were designed for poultry fecal inspection using linear discriminant analysis (LDA). A mathematical model for hyperspectral imaging system was proposed ands its system parameters such as excitation and emission filters were optimally determined by LDA. An alternating scheme was proposed for numerical implementation. The optimal excitation and emission filters of a multispectral imaging system for poultry fecal inspection were determined by LDA. The proposed method is applicable to other agricultural products which are distinguishable by their spectral properties. In general, LDA is solved by a generalized eigenvalue problem since there is no specific constraint. However, only positive functions are meaningful for the excitation and emission filters. This imposes constraints to the maximization problem of discriminability. Nonlinear optimization was directly employed and computationally took longer time. Future studies should include a more efficient algorithm with respect to computational geometry. Physical implementation is also important because of limitation in the excitation and emission filters.

Conflict of Interest The authors have no conflicting financial or other interests.

Acknowledgement This work was partially supported by a grant from the Next-Generation BioGeen 21 Program (No. PJ008055), Rural Development Administration, Republic of Korea. It was also partially supported by Technology Development

270

Program for Agriculture and Forestry, Ministry for Food, Agriculture, Forestry and Fisheries, Republic of Korea.

References Bilgili, S. F. 2001. Poultry meat inspection and grading. CRC, Boca Raton, FL. Chao, K., P. Mehl and Y. R. Chen. 2002. Use of hyper- and multi-spectral imaging for detection of chicken skin tumors, Applied Engineering in Agriculture, 18(1): 113-119. Cho, B., and M. S. Kim. 2007. Study on optimal fluorescence excitation and emission bands for poultry surface inspection, Journal of the Korean Society for Agricultural Machinery, 12(2): 438-441. Duda, R. O., P. E. Hart and D. G. Stork. 2001. Pattern Classification, 2nd ed. New York, N.Y.: Wiley Interscience. Fletcher, J. T. and S. G. Kong. 2003. Principal Component Analysis for Poultry Tumor Inspection using Hyperspectral Fluorescence Imaging, International Joint Conference on Neural Networks, 1: 149-153. Food Safety and Inspection Service, United States Department of Agriculture. 2006. Review of the Pathogen Reduction; Hazard Analysis and Critical Control Point Systems Final Rule Pursuant to Section 610 of the Regulatory Flexibility Act, As Amended. Kim, I., Y. R. Chen, M. S. Kim and S. G. Kong. 2004. Detection of skin tumors on chicken carcasses using hyperspectral fluorescence imaging, Transactions of the ASAE 47(5): 1785-1792. Kim, M. S, A. M. Lefcourt and Y. Chen. 2003. Optimal fluorescence excitation and emission bands for detection of fecal contamination. Journal of food protection. 66(7): 1198-1207. Lawrence, K. C., W. R. Windham, B. Park and R. J. Buhr. 2001. Hyperspectral imaging system for identification of fecal and ingesta contamination on poultry carcasses. ASAE Annual Meeting. OSHA. 1999. Chicken disassembly - Ergonomic considerations. Washington, D.C.: U.S. Department of Labor. Available at: www.osha.gov/SLTC/poultryprocessing. Park, B., K. C. Lawrence, W. R. Windham and R. J. Buhr. 2002. Hyperspectral imaging for detection fecal and ingesta contaminations on poultry carcasses. Transactions of the ASAE 45:2017-2026.