Source: UNIV OF ALABAMA submitted to
DEVELOPING MICROWAVE IMAGING SYSTEM WITH MACHINE LEARNING FOR DETECTING FOREIGN OBJECTS IN PACKAGED FOOD
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
ACTIVE
Funding Source
Reporting Frequency
Annual
Accession No.
1031012
Grant No.
2023-67022-40627
Cumulative Award Amt.
$300,000.00
Proposal No.
2022-11192
Multistate No.
(N/A)
Project Start Date
Jul 15, 2023
Project End Date
Jul 14, 2026
Grant Year
2023
Program Code
[A1521]- Agricultural Engineering
Recipient Organization
UNIV OF ALABAMA
BOX 870344
TUSCALOOSA,AL 35487
Performing Department
(N/A)
Non Technical Summary
Foreign object contamination in foods post a great threat to the safety of consumers and may cause enormous loss to food manufactures due to lawsuits and recalls. These foreign objects may be endogenous to the food, such as pits and bone particles, or exogenous, such as glass, metal, wood, plastic, and stone. According to worldwide data from Horizonscan, there had been one food recall per week that was related to foreign objects found in food in 2018. Currently, food manufacturers usually rely on two types of technologies, i.e., metal detection and/or X-ray inspection, for foreign object detection in food production as their HACCPs. A metal detector can find metals, including ferrous, nonferrous, and stainless steel, which may be splinters from machinery or fractions of broken metal fragments. But metal detectors do not respond to other non-metal foreign objects that may contaminate foods and the metal detector operation requires a metal free zone on the production line. On the other hand, an X-ray detector can be used for a wider range of foreign objects, including metal, stone, bone, and hard plastics. However, X-ray equipment is expensive, regulated, unsafe for workers, and inefficient to detect light and thin bones, and other light foreign materials, such as cherry stones and insects [1]. Furthermore, each food product may present special technological challenges, since the physical properties of food and packages may respond differently to available detection systems.We propose microwave imaging technology (MIT) which offers non-invasive, non-ionizing, hygienic, and contactless sensing methods for packaged food. Due to the low radiation power density, it is also safer to human body than X-ray. Furthermore, compared with the commonly used optical sensing technology, it provides better detection performance due to higher signal penetration through food, resilience to smoke/gas/moisture/glare/liquid, and no need of light. Moreover, thanks to the recent advancement of semiconductor technology, a variety of microwave chipsets has recently been available and revealed excellent sensing performance. In addition, the size and cost of microwave components have dramatically decreased for rapid prototyping and mass production. However, these state-of-the-art microwave chipsets have yet to be utilized in food inspection (e.g., to automate packaged food-processing) which demands higher efficiency in quality assessment and contamination detection. Our research team will measure electromagentic wave propagation through packaged food and construct the corresponding images to detect foreign objects in packaged food. Machine learning algorithms will be developed and used to precisely identify shape, size, material (metal, glass, plastic, bone, etc), and location of the foreign objects inside the packaged food. The proposed microwave imaging system will broadly benefit food-packaging and food-processing facilities.
Animal Health Component
25%
Research Effort Categories
Basic
50%
Applied
25%
Developmental
25%
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
40272102020100%
Goals / Objectives
To fill the research gap and advance food production automation, our goal is to develop non-invasive, non-ionizing, hygienic, contactless, and intelligent in-line food inspection systems empowered by MIMO (Multiple Input Multiple Output) microwave imaging coupled with machine learning algorithm. Specifically, in this proposed project, we propose to design such systems for the detection of foreign objects in packaged food. Specific objectives areObjective 1: Develop a microwave dual-polarized MIMO-based sensing system and machine learning algorithm using simulation models.Objective 2: Assess the performance of the microwave sensing system using real packaged food.The proposed microwave imaging system will include an ultra-wideband (UWB) antenna array in a circular arrangement fashion around an inline packaging facility. The unique features are to support UWB and dual polarization microwave propagation to the packaged food to identify foreign objects that can cause hazard to human. In addition, radar-based image construction with MIMO will expedite the imaging process with high pixel resolution.
Project Methods
Objective 1: Develop a microwave UWB and dual-polarized MIT system and machine learning algorithm using simulation models.Task 1.1 Establish simulation models to optimize UWB and dual-polarized MI systemAmong many packaged foods, the research team will focus on three common cases of foreign objects in packaged food - 1) bone in a ground beef, 2) glass in jelly jar, and 3) metal in peanut butter. The purpose of developing the food models is to establish an EM model that represents the impurities in the packaged food and is used to build the UWB and dual-polarized MIT system. These foods and impurities will be modeled in microwave regime by assigning permittivity and conductivity in the commercial EM full-wave solver. An appropriate operating frequency band will be determined to allow sufficient signal penetration through a packaged food model for image generation. The simulation model will be validated with the previous literature and theory. In addition to the food models and the optimal frequency band, transmitting and receiving antenna arrays will be designed with HFSS. First the geometry of a single antenna element which can operate in UWB frequency for improving image quality. To increase accuracy and precisely locate a foreign object, dual polarization will be supported in the antenna. Then, an array will be constructed with the single elements in a circular fashion. The number of antenna array will be determined to obtain appropriate image quality for machine learning algorithms. The radius of the UWB antenna array will be determined based on image resolution required to identify the location and size of impurity.Task 1.2 Learning enhanced image reconstructionWe will develop a learning-based scheme on the top of conventional image reconstruction method. Briefly, we develop a deep learning network to learn a mapping relationship between image dataset obtained from a low-quality setup, and image dataset obtained a high-quality setup. This deep learning will be based on our previous setup to learn low resolution medical images and high resolution medical image. We will adopt residual channel attention block structure, which has been demonstrated with superior performance in reconstructing under-sampled spatial data. The developed algorithm will serve as an image quality enhancement toolbox. During on-site imaging, we will choose a simplified antenna setup to reconstruct an LQ image. Then, the reconstructed image will be sent to the image enhancement software, to generate a HQ image. Noticeably, the image qualify boosting process only takes around ~25 ms per our preliminary study.Task 1.3 Segmentation-based object analysisIn HQ image, we will conduct image analysis by drawing the boundary of foreign objects (i.e., image segmentation). To draw the boundary of objects, we will assign pixel-level label to all pixels in microwave image. On the basis of existing deep learning framework, U-Net, we will customize our segmentation framework in directions: focal loss and transfer learning. Loss function is a critical design in deep learning framework which decide how to efficiently salient features from existing data. Inspired by our previous work, we will set our focal loss as FL(Pt)=-(1-Pt)rlog(Pt) to address the data imbalance issue, which corresponds the fact that only a small number of pixels in microwave image are related to the object we aim to segmentation. Here, Pt is the ground-truth probability of an object. The hyperparameter r modulates the focusing task. Setting r down-weights the loss for easily classified tissue classes (Pt >0.5), putting more focus on classes hard to classify (i.e., the object). Moreover, to make sure the network design can be quickly adapted in packaged food and poultry detection, the neural network will be optimized in a transfer learning manner. In particular, an incremental learning scheme will be adopted to quickly update the deep learning architecture by incrementally updating new labels and data.Objective 2: Assess the performance of the microwave sensing system using packaged food.Task 2.1 Design and build an automated UWB and dual-polarized MIT sensing systemA prototype of the UWB dual-polarized imaging system will be built in the PI's laboratory. The antennas will be fabricated using laser milling machine to precisely realize the antenna designed in Objective 1. The antennas will be characterized in two anechoic chambers including both ETS Lindgren to and Howland 2100 which the PI has access to at UA. The circular array configuration will be made with a flexi glass which is flexible to bend to provide 360o perimeter to hold the antennas. A low-loss and broadband RF switching network will be incorporated to provide all possible signal path channels from one antenna to the other antennas. A micro-controller will provide digital logic signals to adjust switching sequences to capture EM transmission and reflection. The transmitting EM wave will be generated from a vector network analyzer which can support extremely wide frequency band from 10 MHz to 50 GHz. A computer code will be made to allow automatic and accurate data collection of the packaged food and image generation. We will implement our software in a multi-GPU setup to speed up the analysis for the high volume of data acquired in this task. In particular, we will use a multi-GPU computer, which is currently housed in Dr. Gan's lab, to assign the image reconstruction and image processing to separate GPU(s). Distributed measurements for parallel computing will be used as a special plugin in the algorithms developed in Task 1.2 with the sole aim of speeding up the process. The targeted data collection to imaging time is expected to be less than 30 seconds during the movement on the conveying belt. The packaged food will be purchased from grocery store including Publix, Walmart, and Costco in Alabama.Task 2.2 Validate system and algorithm using packaged food samplesThe performance of the developed microwave-based inspection system and machine learning design will be tested on the real packaged food samples. Different types of packaged food will be purchased from the grocery and may be processed into desired samples for evaluation. Packaged food samples with well-defined bone shape or tissue boundary on surface, such as T-bone steak, chicken wings, and pork belly, will be used on the microwave-based system. Images will be generated through reconstruction from the microwave data and compared with visual images taken using a digital camera. For each type of packaged food sample, we will conduct an initial trial from which the number of microwave images required will be determined by power analysis. All microwave images will be manually annotated by drawing the boundary of each tissue guided by visual images. We will conduct five-fold cross-validation on the dataset. The whole dataset will be randomly split to five groups, each with 20% of the images. We iterate the groups by choosing four groups for training and one group for testing. After five iterations, the mean value of Dice coefficient will be used to evaluate the accuracy of segmentation. Priro to our study, the best performance method in segmenting microwave imaginge is a Dice coefficient of 0.9. We expect the proposed method improve the segementation method and achieve a Dice coefficient of 0.94. Under the cross-validation setup, we will have 2736 images for training and 684 images for validation. With a total of 684 images in each validation, we will have a 99% power to detect if the proposed model is significantly better than existing practice using a one-sided one-sample proportional test at a significance level of 5%.Project output evaluationTask 1.1 & 1.2 for Q1, Q2 and Q3Task 1.3 and 1.4 for Q2, Q3, and Q4Task 2.1 for Q4 and Q5Task 2.2 for Q5 to Q7Manuscripts and reports: Q4 to Q8

Progress 07/15/23 to 07/14/24

Outputs
Target Audience:We have reached a variety of the target audiences through research article publications and attending a symposium. We disseminated the project results by publishing a journal paper to Journal of Agriculture and Food Research. Additionally, the research work was presented at the University Research Symposium. Moreover, the results were shown to undergraduate and graduate students through the regular and newly offered courses at both The University of Alabama and Steven Institute of Technology. We expect our research will broadly impact all the target audience in the world. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided?Two PhD students (one from Stevens and one from UA) were trained in this project. In addition, one undergraduate student from University of Alabama and one undergraduate student from Stevens Institute of Technology were involved in this project. One senior design team of four undergradaute students in department of electrical and computer engineering at University of Alabama was engaged in this project as well. The PhD students presented their work in the internal bi-weekly regular meetings where they exchanged ideas and provided technical comments on the research work. In a classroom environment, the students created a course project about this research topic and presented their work in the classroom. They learned how to collaborate with researchers from different disciplines and institutions, and how to communicate their findings effectively to various audiences. How have the results been disseminated to communities of interest?We have disseminated the result of this project through many channels. First, a journal paper titled "Few Shot Learning for Avocado Maturity Determination from Microwave Images" was published to Journal of Agriculture and Food Researchthis year. Secondly, anotherjounal paper titled "Emerging Non-invasive Microwave and Millimeter-wave Imaging Technologies for Food Inspection" was accepted to Critical Reviews in Food Science and Nutrition (impact factor=10.2), June 2024.Thirdly, a conference presentation was made at the University Research Symposium this April. Fourthly, a conference paper titled "Broadband Miniaturized Vivaldi Antenna for Microwave Imaging System" was accepted to IEEE Antenna and Propagation Symposium. Fifthly,the project results were also presented in related courses such as BME 571 Machine Learning in Biomedical Engineering, ECE 493/593 Engineering Data Analytics, and senior design presentation. Lastly, a poster presentation was given in the ECE Graduate Poster Competition. What do you plan to do during the next reporting period to accomplish the goals?We are planning to accomplish 1) imaging packaged food such as sausage and peanut butter, 2) validating the capability of the current microwave imaging system to locate foreign objects within the food samples, 3) developing a robust machine learning framework to identify the size, type, location of a foreign object, 4) develop intelligent, learning-based image reconstruction method, and 5) write journal and conference manuscripts for dissemination.

Impacts
What was accomplished under these goals? The research team completed two tasks including development of a microwave imaging system and a machine learning algorithm. First, we developed a microwave imaging system which can operate in a wideband frequency ranging from 2 GHz to 8 GHz. This system consists of an array of 10 wideband Vivaldi antennas, a high-speed switching network with single-pole-10-through and single-pole-double-through switches, a microcontroller, a logic circuit, and a vector network analyzer. The switching network is automated by a computer code and takes 35 seconds to complete imaging a food sample. For the beamforming algorithm, Delay-and-Sum (DAS) was used to reconstruct microwave images. Up to today, we have validated the microwave imaging system by scanning and imaging watermelon, avocado, and eggs. Lastly, we produced 56 microwave images with 11 avocados and built a few-shot learning algorithm to determine the days of ripeness. Additionally, we acquired microwave images with a dozen of eggs. Moreover, we developed machine learning algorithm that is based on the convolutional neural network to classify the ripeness of watermelon. Label correction network is in development to address imperfect labeling issue in avocado and egg images. We are developing learning-based image reconstruction method to explore the possibility of reconstructing high-quality image from limited data acquisition.

Publications

  • Type: Journal Articles Status: Awaiting Publication Year Published: 2024 Citation: N. Jeong, Y. Gan, and L. Kong, "Emerging Non-invasive Microwave and Millimeter-wave Imaging Technologies for Food Inspection," Critical Reviews in Food Science and Nutrition, doi: 10.1080/10408398.2024.2364225, June 2024
  • Type: Journal Articles Status: Published Year Published: 2024 Citation: M. Ahmed, H. Mustafa, M. Wu, L. Kong, N. Jeong, and Y. Gan, "Few Shot Learning for Avocado Maturity Determination from Microwave Images," Journal of Agriculture and Food Research, Volume 15, 2024, doi:10.1016/j.jafr.2024.100977
  • Type: Conference Papers and Presentations Status: Accepted Year Published: 2024 Citation: M. I. Kamrul and N. Jeong, "Broadband Miniaturized Vivaldi Antenna for Microwave Imaging System," accepted to IEEE Antenna and Propagation Symposium, July 2024