基于YOLO-DRR与实时处理FPGA边缘计算平台的低空视角下柑橘冠层分割

Implementation of Low-altitude Citrus Canopy Segmentation Based on YOLO-DRR and Real-time FPGA Edge Computing

摘要:

冠层是柑橘果树光合作用的主体，直接影响果树的生长、产量与果实品质，是果树健康与丰产的基础。通过对冠层结构的监测，可以及时调整修剪、灌溉、施肥等种植管理措施，从而优化冠层内部环境，促进果树的健康生长和发育。针对柑橘果园种植密集、冠层之间重叠遮挡影响果树生长效率和产量品质等问题，本研究构建了自然环境下的柑橘果树冠层数据集，提出了YOLO-DRR（YOLO v5s-seg-DSConv-RFEM-RIME）轻量化分割模型，同时为了提高实时性、降低功耗和便于在果园使用，将该模型部署到便携的边缘计算平台中。本研究以YOLO v5s-seg模型为基础，使用多尺度特征提取模块对骨干网络进行改进，提高了多尺度目标的分割精度；使用分布移位卷积模块替换了颈部网络中的C3模块，降低了卷积核中的内存使用量，从而提高了运算速度；采用霜冰优化算法优化YOLO-DRR模型超参数，利用群智能的迭代机制进一步提升模型性能。最后，将YOLO-DRR模型移植部署至FPGA边缘计算平台，利用FPGA强大的边缘计算能力，确保数据处理的实时性，更加有效地利用硬件资源，减少功耗和散热问题，实现复杂背景下柑橘果树冠层长时间实时分割监测的要求。实验结果表明，YOLO-DRR模型对冠层进行分割的精确度达到86.34%，召回率达到88.68%，mAP@0.5达到93.41%，mAP@0.5:0.95达到63.13%，移植至边缘计算平台后，检测速度达到了19f/s，功耗仅为22W，这表明本研究提出的模型具有在复杂背景下对柑橘果树冠层进行实时分割的能力，能够满足果园实时监测冠层生长环境的需求。

Abstract:

The canopy is the primary source of photosynthesis in citrus fruit trees, and it has a direct impact on the growth, yield, and fruit quality of the trees. It is the foundation for healthy and productive fruit trees, and efficient and accurate monitoring of canopy growth is especially important. Monitoring the canopy structure allows planting management measures such as pruning, irrigation, and fertilization to be adjusted promptly, optimizing the internal environment of the canopy and promoting healthy fruit tree growth and development. A dataset of citrus fruit tree canopies in a natural environment was created and a lightweight YOLO-DRR segmentation model (YOLO v5s-seg-DSConv-RFEM-RIME) was proposed to address the issues of dense planting in citrus orchards and overlap shading between canopies, which affect fruit tree growth efficiency and yield quality. Meanwhile, the model was deployed on a portable edge computing platform to improve real-time performance, reduce power consumption, and make it easier to use in inter-orchard scenarios. Firstly, the segmentation accuracy of multi-scale targets was improved based on the YOLO v5s-seg model by using the scale-aware RFE module (RFEM) for backbone networks. Secondly, the use of the distribution shifting convolution module (DSConv) to replace the C3 module in the neck network reduced memory usage in the convolutional kernel, thereby increasing the speed of operations. Thirdly, the rime optimization algorithm (RIME) was used to optimize the hyperparameters of the YOLO-DRR model, and the iterative mechanism of swarm intelligence was utilized to further improve the model performance. Finally, the YOLO-DRR model was transplanted and implemented on the FPGA edge computing platform. The FPGA device was highly environmentally adaptable and can operate reliably in a wide range of temperature and humidity conditions, ensuring the device’s dependability and stability in the complex and changing environment of the orchard. Simultaneously, the powerful edge computing capability of FPGA was used to ensure real-time data processing, more efficient use of hardware resources, reduction of power consumption and heat dissipation issues, and the realization of the requirement for long-term real-time segmentation monitoring of citrus fruit tree canopies in complex environments. The experimental results showed that the YOLO-DRR model segmented the canopy with 86.34% precision, 88.68% recall, 93.41% mAP@0.5, and 63.13% mAP@0.5:0.95. After porting it to the edge computing platform, the detection speed was increased to 19f/s while consuming only 22W of power. This suggested that the model proposed was capable of segmenting the canopy of a citrus fruit tree in the complex context of real-time canopy segmentation, which can meet the demand for real-time monitoring of the canopy growth environment in the orchard.