A 1999 Applied Optics study demonstrated quantitative, nondestructive inspection of internal green mold in whole beans (Phaseolus coccineus L., termed 'soya beans' in the source) by multispectral near-infrared transmission imaging with three lasers (810, 1064, 1319 nm), an eigenvalue-ratio spectral-separability analysis, and least-squares spectral unmixing into per-pixel concentration maps. Only green mold produced a spectral deviation above the measurement noise (relative order 10−310^{-3}); white mold, brown mold, and parching/stiffening fell below it, and the origin of the green-mold spectral signature (a pronounced blueshift and a visible absorption band) was undetermined. The problem is the completing program that follows from that demonstration: determine the chemical/optical origin of the green-mold signature, extend spectrally based detection to the three remaining defect classes and quantify detection limits and unmixing robustness. The problem is open; its significance is local, because the field now pursues the broader sorting goal mainly by reflectance-mode hyperspectral classification.

Background

Whole beans destined for food use must be inspected for internal defects that are invisible from the outside: molds (observed as green, white, and brown colonies in stored beans) and subcutaneous parching or stiffening, a drying and hardening degradation of the inner bean tissue. Conventional industrial practice — specific-gravity separation (floating damaged beans in liquid), visual appearance sorting, and surface spectroscopy at visible wavelengths — acts only on external properties or destroys the sampled beans, so it cannot grade the interior of every bean nondestructively on a processing line. Near-infrared (NIR) radiation, conventionally the 700–2500 nm wavelength band, penetrates seed coats and millimeters to centimeters of plant tissue, which makes transmission imaging a candidate for observing interior anomalies in whole beans.

A 1999 Applied Optics study by Nakamura, Kobayashi, and Kawata built such a system for beans of Phaseolus coccineus L. (the runner bean, although the source calls them 'soya beans', which botanically is Glycine max; the source's usage is retained here). Three continuous-wave NIR lasers at 810, 1064, and 1319 nm illuminated whole beans, and an NIR vidicon camera (a vacuum-tube video camera sensitive in the near infrared) recorded transmitted intensities; after reference-image division and gamma correction, each pixel yields an absorbance image at each wavelength (multispectral transmission imaging). In parallel, diffuse-reflectance spectra — spectra of light scattered back out of an illuminated spot on the bean surface — were measured from 650 to 2500 nm for healthy beans and for beans carrying each defect class. The analysis combined two chemometric ideas (chemometrics being the application of multivariate statistics to chemical and spectral measurement data), following a lineage of diffuse-optics modeling of layered turbid media (Anderson and Parrish, 'The Optics of Human Skin') and eigenvalue-based treatment of NIR spectra (Asker and Kokot): (i) an eigenvalue analysis solving the eigensystem of the spectral covariance matrix [P]t[P]x=ζx[P]^t[P]x=\zeta x built from ten measured spectra (five healthy, five defective), whose ratio of the principal to the second eigenvalue serves as a scalar metric of how strongly a defect class changes the spectra relative to healthy tissue, comparable directly against measurement noise; and (ii) a least-squares spectral-unmixing step, in which measured mixed spectra are modeled as a linear combination of pure-component spectra (healthy tissue and defect tissue) whose estimates are obtained by least-squares fitting in the eigenvector basis. The fitted component spectra supply absorptivities at the three laser wavelengths, and the per-pixel absorbances are inverted under a Beer–Lambert-type linear mixing model (absorbance proportional to absorptivity times concentration times thickness) to produce spatial maps of defect concentration, with the ratio of the defect map to the healthy map cancelling the local bean thickness.

The decisive experimental finding, as reported in the source's analysis, was that only green mold produced a spectral deviation above the measurement noise: the eigenvalue ratios were about 1.95×10−21.95\times10^{-2} for green mold but 7.46×10−47.46\times10^{-4} for white mold, 2.63×10−42.63\times10^{-4} for brown mold, and 8.22×10−48.22\times10^{-4} for parching/stiffening, against an experimental noise level of relative order 10−310^{-3}. The paper therefore demonstrated and quantitatively mapped only the green-mold case, and its discussion left a well-defined but unexecuted program for future studies.

Since 1999 the applied field has developed extensively along a different route: reflectance-mode multispectral imaging (measurements at a few discrete wavelength bands) and hyperspectral imaging (measurements at tens to hundreds of contiguous bands) combined with chemometric or machine-learning classifiers now detects and classifies moldy or fungus-infected seeds of several commodities (pulses, cowpea, soybean) with high classification accuracy, and whole-fruit transmission spectroscopy detects internal mold in apples, tomatoes, and tamarind at classification level. What this later literature does not provide for beans is the specific quantitative program of the 1999 method: per-pixel internal concentration of each defect class recovered from a small set of transmission wavelengths, with quantified minimum detectable concentrations, known robustness of the linear unmixing, and an identified physico-chemical origin of the defect spectra.

Problem Statement

The unresolved problem is to complete the quantitative nondestructive inspection capability that the three-laser NIR transmission method demonstrated only for green mold. Its components are scientifically linked by one objective: extending quantitative internal-defect detection to all defect classes of whole beans with known performance.

  1. Spectral origin. Determine the chemical/optical origin of the pronounced blueshift (an apparent displacement of absorption features toward shorter wavelengths) and of the visible-wavelength absorption band observed in the green-mold spectra relative to healthy bean tissue — that is, identify the mold-derived dye molecule(s) or biochemical modification responsible, by chemical and optical analysis.

  2. Detectability of the remaining defect classes. Establish spectrally resolved detection of white mold, brown mold, and parching/stiffening, whose measured spectral deviations (eigenvalue ratios of order 10−410^{-4}) fell below the original noise level of relative order 10−310^{-3}.

  3. Quantified detection limits and robustness. Quantify the minimum detectable internal concentration of each defect class (limit of detection) and the robustness of the linear spectral unmixing to model mismatch, including spatially varying absorptivities and scattering-induced nonlinearity in the turbid bean tissue.

References

  1. Osamu Nakamura, Makoto Kobayashi, Satoshi Kawata (1999). Nondestructive inspection of Phaseolus coccineus L soya beans by use of near-infrared lasers. Applied Optics 38(12):2724-2727. DOI: 10.1364/ao.38.002724. URL: https://doi.org/10.1364/ao.38.002724
  2. R. Rox Anderson, John A. Parrish (1981). The Optics of Human Skin. Journal of Investigative Dermatology 77:13-19. DOI: 10.1111/1523-1747.ep12479191. URL: https://doi.org/10.1111/1523-1747.ep12479191
  3. N. Asker, S. Kokot (1991). The Application of NIR Spectroscopy for the Prediction of Properties of Australian Refined Reformate. Applied Spectroscopy 45:1153-1157. DOI: 10.1366/0003702914336057. URL: https://doi.org/10.1366/0003702914336057
  4. J.M. Low, W.S. Maughan, S.C. Bee, M.J. Honeywood (2001). Sorting by colour in the food industry. In: Instrumentation and Sensors for the Food Industry, pp. 117-136. DOI: 10.1533/9781855736481.1.117. URL: https://doi.org/10.1533/9781855736481.1.117
  5. S.C. Bee, M.J. Honeywood (2002). Colour sorting for the bulk food industry. In: Colour in Food, pp. 115-142. DOI: 10.1533/9781855736672.1.115. URL: https://doi.org/10.1533/9781855736672.1.115
  6. D.T. Pham, R.J. Alcock (2003). Image Acquisition and Enhancement. In: Smart Inspection Systems, pp. 35-68. DOI: 10.1016/B978-012554157-2/50002-X. URL: https://doi.org/10.1016/B978-012554157-2/50002-X
  7. P. Jannok, K. Petcharaporn, S. Teerachaichayut (2014). Detection of internal mold infection in tomato by transmittance near infrared spectroscopy. Acta Horticulturae 1053:93-99. DOI: 10.17660/actahortic.2014.1053.7. URL: https://doi.org/10.17660/actahortic.2014.1053.7
  8. S. Teerachaichayut, S. Suktanarak, S. Kasemsumram (2014). Non-destructive detection of internal mold infection in sweet tamarind using short wavelength near infrared spectroscopy. Acta Horticulturae 1053:113-119. DOI: 10.17660/actahortic.2014.1053.10. URL: https://doi.org/10.17660/actahortic.2014.1053.10
  9. Souphaphone Phetchalern, Hiroto Tashima, Yuya Ishii, Takeshi Ishiyama, Shinichi Arai, Mitsuo Fukuda (2014). Near-infrared imaging equipment that detects small organic substances in thick foods. Proceedings of SPIE 9192:91921A. DOI: 10.1117/12.2061455. URL: https://doi.org/10.1117/12.2061455
  10. G.N. Hill, K.J. Evans, R.M. Beresford, R.G. Dambergs (2013). Near and Mid-Infrared Spectroscopy for the Quantification of Botrytis Bunch Rot in White Wine Grapes. Journal of Near Infrared Spectroscopy 21:467-475. DOI: 10.1255/jnirs.1083. URL: https://doi.org/10.1255/jnirs.1083
  11. K. Karuppiah, T. Senthilkumar, D.S. Jayas, N.D.G. White (2016). Detection of fungal infection in five different pulses using near-infrared hyperspectral imaging. Journal of Stored Products Research 65:13-18. DOI: 10.1016/j.jspr.2015.11.005. URL: https://doi.org/10.1016/j.jspr.2015.11.005
  12. L. Senni, P. Burrascano, M. Ricci (2016). Multispectral laser imaging for advanced food analysis. Infrared Physics & Technology 77:179-192. DOI: 10.1016/j.infrared.2016.06.001. URL: https://doi.org/10.1016/j.infrared.2016.06.001
  13. Xi Tian, Qingyan Wang, Wenqian Huang, Shuxiang Fan, Jiangbo Li (2020). Online detection of apples with moldy core using the Vis/NIR full-transmittance spectra. Postharvest Biology and Technology 168:111269. DOI: 10.1016/j.postharvbio.2020.111269. URL: https://doi.org/10.1016/j.postharvbio.2020.111269
  14. Carlos Henrique Queiroz Rego, Fabiano França-Silva, Francisco Guilhien Gomes-Junior, Maria Heloisa Duarte de Moraes, André Dantas de Medeiros, Clíssia Barboza da Silva (2020). Using Multispectral Imaging for Detecting Seed-Borne Fungi in Cowpea. Agriculture 10(8):361. DOI: 10.3390/agriculture10080361. URL: https://doi.org/10.3390/agriculture10080361
  15. Xijun Shao, Zhen Guo, Yifei Qin, Jicheng Zhao, Yemin Guo, Xia Sun, Fangling Du (2025). Synergistic multi-level fusion framework of VNIR and SWIR hyperspectral data for soybean fungal contamination detection. Food Chemistry 492:145559. DOI: 10.1016/j.foodchem.2025.145559. URL: https://doi.org/10.1016/j.foodchem.2025.145559
  16. Tadaetsu Hirao, Kazuki Kobayashi, Yasunori Saito (2014). Principle Validation Experiment of Identification Technology of Edible Oil and Inedible Oil Using Laser Diodes. The Review of Laser Engineering 42(6):497. DOI: 10.2184/lsj.42.6_497. URL: https://doi.org/10.2184/lsj.42.6_497
  17. Yizhe Zhang, Jipeng Huang, Qiulei Zhang, Jinwei Liu, Yanli Meng, Yan Yu (2022). Nondestructive determination of SSC in an apple by using a portable near-infrared spectroscopy system. Applied Optics 61(12):3419. DOI: 10.1364/ao.455024. URL: https://doi.org/10.1364/ao.455024