Implementation of an FPGA-based CNN Accelerator for Soybean Seed Quality Analysis

Soybeans are a crucial commodity in global agriculture, and ensuring their quality is essential for maintaining the efficiency and competitiveness of the agricultural industry. In this work, we propose an FPGA-based Convolutional Neural Network (CNN) accelerator for the analysis of soybean seed quality, leveraging the FINN framework to optimize CNNs for deployment on resource-constrained edge devices. Our proposed solution aims to address the high computational and energy demands of CNNs by offloading the inference process to the edge, using the PYNQ-Z1 development board with a Xilinx ZYNQ-7000 APSoC. The implementation explores quantization techniques with 1-bit (1w1a) and 2-bit (2w2a) weights and activations to balance accuracy and throughput. Experimental results show that the 1w1a configuration achieves a classification accuracy of 94%, while the 2w2a configuration improves accuracy to 96%. The results demonstrate that the trade-off leads to up to a 10.65× increase in throughput, with resource usage growing by as much as 86% between the slowest and fastest accelerators in the 1w1a configuration. Additionally, power analysis reveals a 2.4× increase in dynamic power consumption across different configurations.