A hardware-accelerated lightweight CNN for rice pest and disease detection using FPGA and fused knowledge distillation

Rice disease detection constitutes a critical component in the development of intelligent agriculture. Computer vision technology centered on convolutional neural networks (CNN) demonstrates broad application prospects in the agricultural sector. However, CNN models tailored for modern agricultural applications often feature complex architectures and high computational demands, which restrict their deployment on resource-constrained edge devices. For this purpose, a collaborative optimization strategy combining model lightweighting and hardware acceleration is proposed. While optimizing the network structure through strong teacher knowledge distillation (DIST), a hardware acceleration system is designed. The final implementation utilized 186 DSP (Digital Signal Processor) resources, achieving a power consumption of 2.86 W and a throughput of 31.2 GOPS, resulting in an energy efficiency of 10.91 GOPS/W. The classification accuracy for rice pests and diseases reached 93.12%.