Lightweight Network-Based Semantic Segmentation for UAVs and its RISC-V Implementation

作者：Yankai Chen, Hongkun Du, Yutong Zhou · 发表于：Journal of Technology Innovation and Engineering · 年份：2025 · DOI：10.63887/jtie.2025.1.2.9 · 被引用次数：4 · 研究领域：Advanced Neural Network Applications、Adversarial Robustness in Machine Learning、UAV Applications and Optimization

This study proposes a lightweight solution for real-time, low-power semantic segmentation in drone applications based on a RISC-V heterogeneous architecture. A lightweight model integrating U-Net and MobileNetV2 was designed to maintain multi-scale feature extraction capabilities while reducing computational complexity to one-tenth that of standard convolutions through depthwise separable convolutions. Leveraging the modular flexibility of the RISC-V processor and the hardware acceleration capabilities of FPGA, a heterogeneous computing framework was established, supporting customized instructions (e.g., DCONV, TCONV) and on-chip SRAM tiling optimization. Experimental results demonstrate that the model achieves an mIoU of approximately 0.2 and 70% pixel-level accuracy on the validation set, with inference latency reduced by 3–5 times via FPGA acceleration. The TrustZone module ensures secure model deployment through SM3/SM4 cryptographic validation. This work provides a scalable open-source framework for high-precision semantic segmentation on edge devices, validating the engineering feasibility of RISC-V in vision-centric edge intelligence.