基于轻量化DeepLabV3+的作物茎粗测量方法

Crop Stem Diameter Measurement Method Based on Lightweight DeepLabV3+

摘要:

茎粗是作物长势的关键表型参数，其原位实时无损监测对优化农业生产管理具有重要意义。本研究利用集成VPU加速单元的OAK-D相机获取番茄茎秆RGB-D数据，对以MobileNetV2为主干的DeepLabV3+语义分割模型实施剪枝与微调，构建轻量化网络。依托该轻量化网络输出的高保真二值化分割掩膜，利用骨架线算法提取茎秆平面像素宽度，并融合区域深度信息解算真实茎粗。试验结果表明，与原DeepLabV3+、U-Net、PSPNet及BiSeNetV2模型相比，轻量化DeepLabV3+展现出卓越的工程部署优势。参数量减至1.48×10^6，模型内存占用量仅4.03 MB，在OAK-D端帧率为5.50 f/s。在参数规模大幅缩减的同时，模型像素级感知力并未退化，精确率、召回率、F1分数与IoU分别达96.03%、96.19%、96.11%与92.51%。基于该高质量分割基底，茎粗计算值均值与中值决定系数R2分别达到0.987和0.990。所提面向边缘设备的轻量化测量方法，可为温室移动在线巡检装备的嵌入式感知系统设计及智慧灌溉管理提供理论依据与技术支撑。

Abstract:

Stem diameter is a critical phenotypic parameter of crop growth. Its in-situ, real-time, and non-destructive monitoring is of great significance for optimizing agricultural production management. It utilized an OAK-D camera integrated with a VPU acceleration unit to acquire RGB-D data of tomato stems. Pruning and fine-tuning were utilized to a MobileNetV2-based DeepLabV3+ semantic segmentation model to construct a lightweight network. Relying on the high-fidelity binarized segmentation masks outputed by this lightweight network, a skeletonization algorithm extracted the planar pixel width of the stems. This width was then fused with regional depth information to calculate the actual stem diameter. Experimental results indicated that, compared with the original DeepLabV3+, U-Net, PSPNet, and BiSeNetV2 models, the lightweight DeepLabV3+ demonstrated remarkable engineering deployment advantages. Its parameter count was reduced to 1.48×10^6, the model memory footprint was only 4.03 MB, and the frame rate on the OAK-D side was 5.50 f/s. Despite the substantial reduction in parameter scale, the pixel-level perception capability of the model did not degrade. Precision, recall, F1 score, and IoU reached high levels of 96.03%, 96.19%, 96.11%, and 92.51%, respectively. Based on this high-quality segmentation foundation, the mean and median determination coefficients (R2) of the calculated stem diameters reached 0.987 and 0.990, respectively. The proposed lightweight measurement method oriented toward edge devices provided a theoretical basis and technical support for the design of embedded perception systems in greenhouse mobile inspection equipment and smart irrigation management.