Progress 02/16/24 to 02/15/25
Outputs
Target Audience:Our primary audience is the small and medium-sized meat and poultry processors (MPP) in the U.S. Our project partner, Lorentz Meats of Cannon Falls, MN belongs to that category. We estimate that this primary target audience would benefit most from the meat monitoring and lean estimation tool that we are developing using image analysis and artificial intelligence/machine learning (AI/ML). Our near-term MPP target audience will include individual organizations, cooperatives and associations that will benefit from the project. Our longer-term goal will be to expand this project to support other food processing organizations where appropriate. Most of our target audiences in the small to medium processor market segment are in communities with population groups who are socially, economically, or educationally disadvantaged. We attended the 85th Convention of American Association of Meat Processors (AAMP) in August 2024 to reach out to a larger group in this target audience. We are also working with our commercialization partner, Larta Institute, to gain access to our target audience. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?Training Activities: We hired a summer intern, Arijit Ghoshal, an undergraduate student at University of California, Berkeley pursuing dual degrees in mathematics and computer science with an academic background in machine learning to work on this project. He made valuable contributions over the summer in devising and implementing convolutional neural network models for image processing and lean estimation. He has continued to work on a part-time basis after returning to school as a junior. Mr. Ghoshal has been fully supported by SC's internal R&D funds. We believe that Mr. Ghoshal has substantially benefited from this opportunity to perform R&D and product development work under the guidance of senior engineers at SC. Professional development: The work on this project is helping to expand the breadth of knowledge of our team in digital image processing and analysis, and AI/ML techniques. The project has helped us expand our professional expertise into the food industry in general and meat processing in particular. Two members of the team attended the 85th American Convention of Meat Processors & Suppliers' Exhibition in Omaha, NB in August 1-3, 2024, for customer discovery and gaining a better understanding of the industry. How have the results been disseminated to communities of interest?
Nothing Reported
What do you plan to do during the next reporting period to accomplish the goals?The following is a list of our proposed activities for the no-cost extension period: Continue improving and testing the three tool prototypes currently operating at Lorentz Meats. We are in the process of replacing the tablet (used to display the real-time lean estimation to the sorter) with a small flat-panel display. We will first make this change on one of the units and test it for a month before replacing the tablets in the other two prototypes. Implement the commercialization strategy being developed with the help of the team at Larta Institute. These activities include implementing marketing strategies such as developing a website for the product and creating brochures. Additionally, we will develop a pricing strategy. Explore opportunities for additional funding for productization of the prototype. Reach out to other small and medium-sized meat processors to find one or more partners to test the prototype in a different environment. Continue refining the lean estimation algorithms with the large amount of new data (about ten new data points per day) to help improve the accuracy of the estimate. Implement a secondary functionality to the prototype by leveraging the AI/ML technology that we have deployed in order to make the tool more versatile and more viable as a product. This functionality is the detection of foreign objects like knives or gloves that may have accidentally fallen into the bin and partly hidden by the meat.
Impacts
What was accomplished under these goals?
Problem addressed by this project: Meat and meat products have become an increasingly important part of the daily diet for people in the U.S. and around the world. With greater demand has come a need for greater production efficiency in the face of worker shortages, a problem highlighted during the Covid-19 pandemic. Visual inspection is an important part of meat processing for ensuring quality and food safety. However, visual inspection, e.g., for fat content estimation, is approximate and may contain meat with a wide range of fat content. The ground meat from each batch is then tested with core samples and adjusted to reach the desired lean-to-fat ratio, leading to delays, stoppage of work and wastage. The off-line inspection tools are also expensive. This project addresses the challenge of improving inspection efficiency for fat content estimation using technology that is affordable by small and medium-sized meat processors. The ultimate goal is to improve quality, productivity and assure compliance with USDA regulatory requirements for meat processors while reducing wastage, using technology that is inexpensive and can be integrated seamlessly into small and medium-size meat processing facilities leading to higher profitability. Immediate beneficiary of this work, and how this benefit is seen: The immediate beneficiary of this work will be Lorentz Meats by improving their productivity, efficiency, and profitability. Subsequently, all small and medium-sized meat processors will potentially be helped by the meat inspection product that we are developing. Progress with goals and objectives listed in project initiation form: We have made significant progress with all the objectives that we listed in the project initiation report. Here, we describe the progress with the list in the same order as the objectives. A suite of algorithms has been developed and implemented in software in the past reporting period. Most of these algorithms are being used by the tool prototype for real-time lean estimation where we are using both conventional regression-based analysis as well as artificial intelligence and machine learning and (AI/ML) for lean/fat ratio estimation. The prototypes that are currently operating are also helping with systematic acquisition of data that will be used to further refine and train our algorithms during the so that we can continue to improve lean estimation accuracy to well within ±3% accuracy. All components of the imaging system are commercial off-the-shelf (COTS). Through careful choices of components, we have been able to bring down the cost of the hardware components in the current version of the prototype is well below our original goal of $10K. This accomplishment will help market the unit at a cost significantly lower than products developed with competing technologies. We have fully met this goal of operating the prototypes for lean estimation on the processing floor as the meat is being sorted into bins without any interruption or any sort of disruption to the work process. The tool protypes currently installed at Lorentz Meats have a small footprint in the form of a mounting frame attached to the processing tables and the wall so that it does not use any floor space. Each stand-alone prototype consists of a box that houses the camera, computer and power supplies. Each unit is deployed using a coat-hangar type of attachment so that it simply hangs from the frame over each bin with the camera looking directly down into the bin through a small window in the box. The prototypes are always powered up and are left in place during the daily cleaning process. We have continued to work closely with our partner, Lorentz Meats, throughout the design, implementation, and testing process. They are fully invested in working with SC to build a system that they can use. The potential for increased productivity and therefore for increased profits (while meeting USDA regulatory requirements) will become evident over the current reporting period, i.e., in the no-cost extension year of the program. SC is working with Larta Instituteto develop a test protocol, milestone objectives, measurable goals, and the business model applicable to SC's future customers in the small to medium-sized meat processors. Key outcomes or other accomplishments realized: The most important accomplishment in the past reporting period was installing three operational prototypes at Lorentz Meats. Multiple iterations in hardware design and choice of components were tested over an eight month-period before we settled on a design that was sufficiently robust to the harsh processing floor environment yet low in cost. The second important accomplishment was the development and training/calibration of algorithms for lean estimation and their software implementation in the Python programming language. Most of our effort was spent in this area. We are currently able to predict lean content with an accuracy of about ±3% of the measured content when the prediction is averaged over a significant number of test samples (30 or more). However, we intend to improve the accuracy during the no-cost extension period. The third accomplishment was developing and implementing a suite of algorithms for other key functionalities that are required in the lean estimation pipeline. One important algorithm that we have implemented is the automatic detection of a change in the bin which is in the view of the camera. Accurate detection of a new bin is important for real-time lean estimation. Other algorithms that were implemented include ML model for selection of specific frames from the video to color analysis for image segmentation to distinguish meat from fat. The fourth important accomplishment was streamlining the data collection process at Lorentz Meats. To that end, we made several modifications to the software, e.g., to enable automatic start and stop of data acquisition. We continually update the software remotely as new versions are available. The fifth important accomplishment was in the commercialization area where we completed a market survey, worked on customer discovery and attended the American Association of Meat Processors (AAMP) annual convention to gain a better understanding of the industry's needs on real-time meat monitoring.
Publications
|
Progress 02/16/23 to 02/15/24
Outputs
Target Audience:Our primary audience is the small and medium-sized meat and poultry processors (MPP) in the U.S. Our project partner, Lorentz Meats of Cannon Falls, MN belongs to that category. We estimate that this primary target audience would benefit the most from the image analysis tool we are developing. Our near-term MPP target audiences will include individual organizations, cooperatives and associations that will benefit from the project. One way to reach out to a larger group in this target audience would be to attend a trade show, and we are considering attending one this summer. Our longer-term goal will be to expand this project to support other food processing organizations where appropriate. Most of our target audiences in the small to medium processor market segment are located in communities with population groups such as racial and ethnic minorities and those who are socially, economically, or educationally disadvantaged. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?Training Activities: We hired a summer intern, Sesha Charla, a graduate student at Purdue University in the doctoral program, to do a literature search for relevant image processing algorithms, evaluate them for this application, and implement some of them in software. We believe that Mr. Charla benefited from this opportunity to perform R&D work under the guidance of senior engineers at SC. Professional development: The work on this project is helping to expand the breadth of knowledge of our team in the area of digital image processing, and remote video data acquisition and analysis. How have the results been disseminated to communities of interest?
Nothing Reported
What do you plan to do during the next reporting period to accomplish the goals?The following is a list of our proposed activities for the second reporting period: Streamline the data collection process currently underway at Lorentz Meats: We will make some minor modification to the hardware (cables of appropriate lengths and with easily distinguishable end connectors), and software (automatic start of data acquisition on computer start up, continually updated software as new versions are available). Modify the current image analysis algorithms to make them more robust to changes in image properties such as brightness, contrast, and saturation as well as changes in the operating environment (e.g., presence of other objects outside the bin that is captured by the camera such as part of a different bin). Train the lean/fat estimation algorithm with the large amount of new data (about ten to fifteen new data points per day) that is becoming available. This large amount of data will help significantly improve the accuracy of the estimate. While we are currently using linear regression for lean/fat estimation, we will extend our approach to include machine learning techniques. Design and implement the first product prototype of this tool. It will be a stand-alone unit hung on the frame above the bin and will consist of two cameras (one looking down at the bin, and one looking at the edge of the conveyor carrying the meat), computer, and battery to power the unit for 4-8 hours of operation. We intend to deliver the first prototype later this year to Lorentz Meats for testing. As part of our effort to disseminate to the meat processing community, we intend to attend the 85th American Convention of Meat Processors & Suppliers' Exhibition at Omaha Nebraska, August 1-3. We will describe our product prototype to gauge interest and get feedback regarding increasing the tool's usefulness to the small and medium-sized meat processors. We will continue to work closely with Larta Institute to understand and quantify the market opportunity, and to develop a commercialization strategy.
Impacts
What was accomplished under these goals?
Problem addressed by this project: Meat and meat products have become an increasingly important part of the daily diet for people in the U.S. and around the world. With greater demand has come a need for greater production efficiency in the face of worker shortages, a problem highlighted during the Covid-19 pandemic. Visual inspection is an important part of meat processing for ensuring quality and food safety. However, visual inspection, e.g., for fat content estimation, is approximate and may contain meat with a wide range of fat content. The ground meat from each batch is then tested with core samples and adjusted to reach the desired lean-to-fat ratio, leading to delays, stoppage of work and wastage. The off-line inspection tools are also expensive. This project addresses the challenge of improving inspection efficiency for fat content estimation using technology that is affordable by small and medium-sized meat processors. The ultimate goal is to improve quality, productivity and assure compliance with USDA regulatory requirements for meat processors while reducing wastage, using technology that is inexpensive and can be integrated seamlessly in small and medium-size meat processing facilities leading to higher profitability. Progress withgoals and objectives listed in project initiation form: We have made significant progress in all the objectives that we listed in the project initiation report. Here, we describe the progress with the list in the same order as the objectives. A suite of algorithms has been developed in the past reporting period. Some of the algorithms are being used for processing the recorded videos and images (extracted from the videos) in preparation for lean-to-fat ratio estimation. We are currently using regression-based analysis for the lean/fat estimation. The systematic acquisition of data that was started in February 2024 will help us calibrate our algorithms so that we can estimate lean/fat ratio to within 3% accuracy. All components of the imaging system are commercial off-the-shelf (COTS). Through careful choices of components, we have been able to bring down the cost of the hardware components in the current version of the prototype to below $10K). We have fully met this goal of running the image processing and analysis software in estimation as the meat is being sorted into bins with no interruption or any sort of disruption to the work process. The tool protype currently installed at Lorentz Meats has a small footprint in the form of a mounting frame attached to the processing tables and the wall so that it does not use any floor space. Currently, it takes less than 20 minutes to install the system in the morning and less than 10 minutes to take it down at the end of the day before the cutting room is cleaned. Over the next month, the deployment will be reduced by half by using proper length of cabling and use of matching colors for connectors. Our goal is to have a final prototype that may be seamlessly deployed using a coat-hangar type of attachment so that the integrated camera and computer unit may simply be hung from the frame over each bin. We are working closely with our partner, Lorentz Meats, throughout the design, implementation, and testing process. They are fully invested in working with SC to build a system that they can use. The potential increased productivity and therefore profits while meeting USDA regulatory requirements will become evident over the current reporting period, i.e., in the second year of the program. SC is working with Larta to develop a test protocol, milestone objectives, measurable goals, and the business model applicable to SC's future customers in the small to medium-sized meat processors. Key outcomes or other accomplishments realized: A key accomplishment in the past reporting period was defining all the details of the problem being addressed. The goal of the program has been to use image analysis for real-time, in situ estimation of the lean/fat ratio. We started with the intent to estimate fat on individual trim pieces as they moved on the conveyor. However, Lorentz Meat is interested in the average lean/fat content of each 2000 lb meat-filled bin (combos) - not the individual trim pieces. If the tool is able to give a running estimation of the fat content of each bin as it is being filled, the sorter can adjust the pieces being added so that the target ratio is attained. In this way, any expensive re-work is avoided because Lorentz guarantees the lean/fat ratio of each combo bin. Since the average fat content of each combo bin is currently measured, we can acquire data to train the algorithms without any extra work on the part of Lorentz. The second important accomplishment was the development of software for all aspects of the imaging system ranging from selection of specific frames from the video to background removal and finally image segmentation analysis for distinguishing meat from fat. The software was implemented in the Python programming language. Most of our effort was spent in this area. The third important accomplishment of this reporting period was the selection and acquisition of the appropriate imaging hardware and the small-footprint computer, and the design of the camera support system and its implementation by Lorentz Meats. Finally, in January and early February, we integrated the hardware and software into the first iteration of the prototype image analysis system for installation at Lorentz Meats. While the installation actually happened during our visit to Lorentz Meats in the first week of the second reporting period, we are pleased to report that the system is now acquiring data reliably every day, and is expected to continue to do so for the next few months. Immediatebeneficiary of thiswork, and how this benefit is seen: The immediate beneficiary of this work will be Lorentz Meats by improving their productivity, efficiency, and profitability. Subsequently, all small and medium-sized meat processors will potentially be helped by the product we are developing.
Publications
|