An FPGA-Integrated System with Attention Mechanism-Based Learning for Artificial Fragrant Rice Identification

Highlights Portable FPGA system for real-time AFR identification, improving food safety. Attention-based model reduces interference and enhances AFR identification accuracy. FPGA implementation with optimizations boosts inference speed by 2.51× and energy efficiency by 1.72×. Abstract. This article presents a portable FPGA-based system for real-time, on-site identification of artificially fragrant rice (AFR), which poses potential health risks. It addresses the limitations of traditional laboratory methods, which are often costly and time-consuming. An attention mechanism-based model is developed for accurate AFR identification, reducing interference from impurities and ensuring reliable results. A top-k engine and an approximate multiplier are designed to optimize the model for FPGA deployment, enhancing computational efficiency and minimizing resource consumption. The customized FPGA architecture achieves 90.5% accuracy in AFR identification, with a 2.51× increase in speed and a 1.72× improvement in energy efficiency compared to typical edge platforms. This research pioneers an efficient, on-site solution for AFR identification, contributing to rapid food safety assessment advancements. Keywords: Artificially fragrant rice (AFR), Attention mechanism, Food safety, FPGA, On-site identification.