A high-efficiency lettuce quality detection system based on FPGA

In recent years, hyperspectral imaging technology has been widely applied in agriculture; however, its complex and redundant data poses significant challenges for implementation on edge computing platforms. To address this issue, this study proposes a high-efficiency lettuce quality detection system based on FPGA, employing a lightweight Lit-FasterNet model. The system performs inference on dimension-reduced hyperspectral images on an FPGA platform to detect nitrogen and potassium quality in lettuce. First, hyperspectral data is subjected to feature dimensionality reduction to extract channels relevant to lettuce quality. Next, targeted structural improvements are made to the FasterNet model, including the addition of fully convolutional layers, optimization of partial convolution channels, and modifications to the fully connected layers, resulting in an efficient LitFasterNet model suitable for lettuce quality detection. Finally, the improved model is mapped onto an FPGA, with DSP and data synchronization strategies employed to optimize resource utilization and efficiency. Dual transmission of image data and weights is achieved using DDR3 SDRAM. Experimental results show that the Virtex-7 platform achieved a classification accuracy of 91.87 %, with a portability accuracy of 98.8 %. The detection speed for a single hyperspectral image is 0.083 s, with a power consumption of only 6.323 W. This study offers a feasible and efficient solution for integrating monitoring and management of crops in large-scale agricultural production, demonstrating significant application value and potential for widespread adoption.