KEYNOTE: Plastic Contamination Research: From Inception to AI-Enhanced Commercial Product

This keynote presents a comprehensive journey from early-stage research to the commercialization of a very low-cost traditional machine-vision color sorter "Plastic Inspection, Detection, and Ejection" system, that was later augmented with AI to form an advanced plastic inspection system. Initially rooted in traditional machine-vision color sorting, the project evolved through the integration of AI, addressing one of the field's most significant challenges: the labor-intensive training of vision models.

Overcoming the necessity for millions of manually annotated images, our approach leveraged general purpose vision-transformer-to-caption models to enable semi-automated image annotation, to create a massive data-set automatically. That was then used to train a high-speed vision transformer model, with exemplary classification performance. The AI addition enabled a black-box solution that emphasizes ease of use, requiring only basic functionality checks for cameras and computers, and eliminates the need for external expert intervention. The presentation critically examines the technical challenges encountered, the innovative solutions implemented, and the alignment of our final product with the original vision for a robust, AI-enhanced commercial tool ready for industrial deployment.

About the presenter
Mathew G. Pelletier, Ph.D., is senior scientist with the U.S. Department of Agriculture’s Agricultural Research Service, where he leads cutting-edge machine-vision and AI initiatives for agricultural engineering. He earned his doctorate in Engineering from the University of California, Davis (1998), specializing in signal processing, computer vision and machine learning. Over the past 25 years he has conceived and delivered a succession of award-winning solutions; from a high-speed, deep-learning system that removes plastic contamination from cotton (that costs U.S. growers over $750 million annually, which resulted in winning two Federal Laboratory Consortium awards; where competition included NASA, NIST and NOAA researchers) to patented real-time trash-detection algorithms and novel microwave and 3D-imaging methods for moisture and material analysis. A prolific inventor, Matthew holds seven U.S. patents for enhanced sensors based on machine learning-vision and Ai deep learning applications.

Pelletier combines technical leadership with extensive management and editorial experience. He has served as project manager, team leader and mentor for multidisciplinary research and development teams for more than two decades; and is the founding chief editor of the AgriEngineering Scientific Journal; served as the associate editor for Engineering in Journal of Cotton Science, and he has authored over a 100 scientific papers and delivered numerous talks at international conferences. His pioneering work in development of novel low-cost embedded sensors has set new benchmarks in the cost constrained industrial ag-tech sector for performance and rapid technology transfer to industry partners