Detection of foreign materials on broiler breast meat using fusion of visible near-infrared and short-wave infrared hyperspectral imaging

Authors: CHUNG, SOO - Oak Ridge Institute For Science And Education (ORISE); Yoon, Seung-Chul

Submitted to: Applied Sciences
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/14/2021
Publication Date: 12/16/2021
Citation: Chung, S., Yoon, S.C. 2021. Detection of foreign materials on broiler breast meat using fusion of visible near-infrared and short-wave infrared hyperspectral imaging. Applied Sciences. https://doi.org/10.3390/app112411987.
DOI: https://doi.org/10.3390/app112411987

Interpretive Summary: During poultry processing, various types and sizes of foreign materials can be unintentionally inserted into chicken meat products. The foreign materials such as pieces of metal and plastic equipment, rubber gloves, and pallets, just to name a few, are physical hazards causing health problems to consumers. The foreign materials are also a concern to the poultry industry because product recalls associated with foreign materials found in final products have caused significant economic loss to the industry. A USDA study was conducted to investigate whether a non-destructive and non-contact imaging technology, called hyperspectral imaging, is feasible to detect surface foreign materials found in unpackaged fresh chicken breast fillets during poultry processing. Hyperspectral imaging is a technology that can extract and process spectral fingerprints of agricultural and food products at each image pixel. The study hypothesized that foreign materials would have different spectral fingerprints compared to the spectra of chicken meat in terms of chemical compositions. Hyperspectral images of fresh 12 fillet samples with and without foreign materials were acquired using two hyperspectral imaging systems operating in two non-overlapping spectral ranges in the visible and near-infrared (400-1,000 nm) and the short-wave infrared (1,000-2,500 nm) wavelength ranges, respectively. The study found that data fusion in the entire 400-2,500 nm wavelength range was more effective than either hyperspectral imaging system alone. The detection accuracy for 5 x 5 mm of foreign materials was about 95%, while the accuracy for detection of smaller 2 x 2 mm of foreign materials was about 91%. The study results suggested that prediction models utilizing spectral features carried in the 400-2,500 nm wavelength range would improve the detection of foreign materials more effectively than using a limited spectral range. The study results also suggested that data fusion of two hyperspectral imaging systems would be feasible for the detection of foreign materials. Once fully developed as an online sorting technology, this technology will contribute to enhance the food safety of chicken meat products.

Technical Abstract: Foreign material (FM) is a physical contaminant that could be unintentionally added to food during processing, and thus reduces the quality and safety of the product. We investigated the image fusion of two pushbroom hyperspectral imaging modalities in the visible-near infrared (VNIR) range of 400–1000 nm and the short-wave infrared (SWIR) range of 1000–2500 nm for detection of FMs on the surface of chicken breast fillets. Thirty different types of FMs that could be commonly found in poultry processing plants were used as sample materials in this study. The FM samples in two different sizes (approximately, 5 x 5 mm and 2 x 2 mm) were prepared and put on the surface of the chicken fillet for hyperspectral image measurements. For the pixel-level classification, Mahalanobis distance metric (full spectra) and key wavebands (multi spectra) applied together to develop discrimination model to detect FMs from chicken fillet. Selected key wavebands were obtained 1) based on the correlation coefficient from partial least square regressing (PLSR) and 2) well-known wavelengths that are related to the ingredients of meat. There was no false detection of FM in the meat image without FM and the detection rate in training set with bigger size of FM (5 x 5 mm) using rule-based decision fusion model was 94.6% and 90.9% in the test set with smaller size of FM (2 x 2 mm). In addition, Support vector machine (SVM) were performed to classify the type of the found FMs at the blob-level. SWIR range-based result showed better classification performance than VNIR range-based result and showed 77.7% (training)-82.5% (test) classification accuracies. This study indicated that fusion of multiple hyperspectral imaging system could detect and classify FMs more effectively than using sole system.