The food service industry must keep premises clean and free of foodborne pathogens that can be harbored in biofilms and organic residues. These may cause foodborne infections, endangering consumers and service providers. New fluorescence technology with advanced artificial intelligence algorithms can be a solution for detecting invisible contamination problems. However, improving algorithms requires access to data, raising concerns about data privacy and potential leaks of sensitive data. We present federated learning, a decentralized privacy-preserving method, to train algorithms for precisely detecting contamination in food preparation facilities and improving cleanliness while providing data privacy assurance for clients in the food-service industry.
This paper introduces an autonomous robot system with an intelligent contaminant detection and disinfection device. The system can maneuver and using a robotic arm can detect, disinfect, and document invisible organic contamination on surfaces that may host pathogens. We will present repeated autonomous detection of hard-to-see potato starch biofilms on a conveyor belt surface. The system will be designed to report the time and location of the detected contamination on a digital floor plan. The system also records the amount of germicidal energy dosage to the surface by calculating the optical power, exposure time and distance to the surface.
Sanitation inspection is an ongoing concern for food distributors, restaurant owners, caterers, and others who handle and serve food. They must prevent food contamination but now must also deal with potential infection spread among workers and customers. Beyond zero tolerance legal requirements and damage to institutional or restaurant reputation, loss of trust with workers and customers can be very costly. We provide fluorescence imaging results that were measured, analyzed, and recorded on different high touch surfaces in restaurants and institutional kitchens. We have developed an algorithm to classify cleanliness levels based on the extent of organic residues detected.
Meat and poultry can be contaminated by pathogens like E. coli and salmonella. Animal fecal matter and ingesta host these pathogens, so developing a method to detect contamination on meat surfaces is crucial. We visited four meat processing facilities and used a handheld fluorescence imaging device to detect fecal matter or ingesta on carcasses. We investigated the efficiency and reliability of a state-of-the-art semantic segmentation algorithm to segment fecal or ingesta contaminated regions in meat surfaces images. Industry could use CSI-D to train employees in trimming carcasses as part of their Hazard Analysis Critical Control Point zero tolerance plan.
Poultry meat is the most consumed meat in the US. To ensure a wholesome and safe product, carcasses and the viscera are inspected for disease and other conditions indicating they should be condemned as unfit for consumption. Septicemia-Toxemia (SepTox) is the most common carcass condemnation observed and reported. In this paper, we present a fast, convenient, and easy-to-use handheld system to detect SepTox for condemnation in post-eviscerated poultry carcasses. We provide fluorescence imaging measurements and analysis on poultry carcasses for developing machine learning models to classify carcasses with SepTox for future high speed process line automated imaging.
Our goal is to use multiple spectroscopy methods in a single system and develop novel multimode spectroscopic data fusion techniques for fish species identification in real-time. We collected spectral signatures of fish fillets from six fish species using four hyperspectral imaging systems: (1) Reflectance spectral imaging in the visible and NIR (VIS-NIR), (2) Reflectance spectral imaging in the short wave infrared (SWIR), (3) Fluorescence visible spectral imaging with UVA and violet excitation, (4) Raman imaging with a 785 nm laser excitation. All fish fillet samples were confirmed by DNA testing. We built multiple classification/ dimension reduction combination methods to calculate the average sensitivity and associated variance for each class and each spectroscopy mode. In our prototype, the derived statistics are used to form policies for Monte Carlo prediction reinforcement learning. We compared the results of our weighted fusion decisions against individual spectroscopy mode decisions to show an overall sensitivity improvement. We believe this is the first reported use of reinforcement learning applied to multimode spectroscopy data classification in food fraud applications.
Our goal is to develop a reliable and cost-effective spectral imaging system with sparse spectral measurements. Relative to standard RGB imaging systems, hyperspectral imaging systems offer superior capabilities but tend to be expensive and complex, requiring either a mechanically complex push-broom line scanning method, a tunable filter, or a large set of LEDs to collect images in multiple wavelengths. We would like to overcome these limitations by employing a novel spectral reconstruction algorithm to recreate the full-resolution reflectance or fluorescence spectrum from an optimized selection of images at a sparse set of wavelengths. This algorithm is aided by a single full-resolution spectrometer measurement representing an average value over the selected spatial scene. We use a genetic algorithm-based methodology to identify the optimal wavelengths for sparse spectral measurement and invoke a cost function that includes a weight vector to emphasize minimization of errors in key portions of the spectrum. To validate the proposed algorithm, reflectance spectra in the visible and NIR (400-1000 nm) and fluorescence spectra with UV illumination were collected from fish fillets to validate our methods. In this paper, we discuss the reconstruction algorithm and the genetic algorithm-based optimization method that we use to determine the optimal set of wavelengths for imagery collection. We also present results from a fish species classification study using the reconstructed spectra as feature sets for four common machine learning algorithms. The classification accuracies based on these reconstructed spectra are on par with the accuracies that result from using the original full spectral resolution data.
This study developed multimode hyperspectral imaging techniques to detect substitution and mislabeling of fish fillets. Line-scan hyperspectral images were collected from fish fillets in four modes, including reflectance in visible and nearinfrared (VNIR) region, fluorescence by 365 nm UV excitation, reflectance in short-wave infrared (SWIR) region, and Raman by 785 nm laser excitation. Fish fillets of six species (i.e., red snapper, vermilion snapper, Malabar snapper, summer flounder, white bass, and tilapia) were used for species differentiation and frozen-thawed red snapper fillets were used for freshness evaluation. A total of 24 machine learning classifiers were used for fish species and freshness classifications using four types of spectral data in three different subsets (i.e., full spectra, first ten components of principal component analysis, and bands selected by a sequential feature selection method). The highest accuracies were achieved at 100% using full VNIR reflectance spectra for the species classification and 99.9% using full SWIR reflectance spectra for the freshness classification. The VNIR reflectance mode gave an overall best performance for both species and freshness inspection.
In this study we develop a methodology to accurately extract and visualize cardiac microstructure from experimental
Diffusion Tensor (DT) data. First, a test model was constructed using an image-based model generation technique on
Diffusion Tensor Magnetic Resonance Imaging (DT-MRI) data. These images were derived from a dataset having
122x122x500 um3 voxel resolution. De-noising and image enhancement was applied to this high-resolution dataset to
clearly define anatomical boundaries within the images. The myocardial tissue was segmented from structural images
using edge detection, region growing, and level set thresholding. The primary eigenvector of the diffusion tensor for each
voxel, which represents the longitudinal direction of the fiber, was calculated to generate a vector field. Then an
advanced locally regularizing nonlinear anisotropic filter, termed Perona-Malik (PEM), was used to regularize this vector
field to eliminate imaging artifacts inherent to DT-MRI from volume averaging of the tissue with the surrounding
medium. Finally, the vector field was streamlined to visualize fibers within the segmented myocardial tissue to compare
the results with unfiltered data. With this technique, we were able to recover locally regularized (homogenized) fibers
with a high accuracy by applying the PEM regularization technique, particularly on anatomical surfaces where imaging
artifacts were most apparent. This approach not only aides in the visualization of noisy complex 3D vector fields
obtained from DT-MRI, but also eliminates volume averaging artifacts to provide a realistic cardiac microstructure for
use in electrophysiological modeling studies.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.