基于改进YOLO 11n的甘薯苗与杂草识别轻量化方法

Lightweight Method for Sweet Potato Seedling and Weed Identification Based on Improved YOLO 11n

摘要:

针对甘薯大田种植复杂环境下小目标和重叠遮挡区域的杂草难以识别、移动设备算力有限实时性较差等问题，本研究对YOLO 11n模型进行改进，提出了一种兼顾精度和速度的甘薯苗与杂草识别轻量化模型（SCSF-YOLO 11n）。首先，采取ShuffleNetV2轻量化网络作为主干特征提取网络，降低模型参数量和浮点运算量;其次，使用Context Guided Block替换颈部网络C3k2模块的瓶颈（Bottleneck）结构，整合目标区域和周边环境信息，减少田间背景干扰;为解决甘薯苗与杂草重叠遮挡导致局部细节特征丢失的问题，将SEAM（Separated and enhancement attention module）注意力机制集成至头部网络，进行多尺度特征提取与融合，提升模型遮挡场景下识别能力;最后，采用Focaler-MPDIoU作为损失函数，加快模型网络收敛速度，降低其损失值。试验结果表明，改进后模型在验证集上的精确率、召回率、平均精度均值（mAP50）分别为93.6%、94.3%、95.4%，参数量仅为1.4×10^6，模型内存占用量为3.2 MB，浮点运算量为3.1×10^9，处理速度为112.6 f/s;与YOLO v5n、YOLOX-tiny、YOLO v7-tiny、YOLO v8n、YOLO v9t、YOLO v10n、YOLO 12n、DINO等模型相比，mAP50分别提升3.0、4.1、3.7、4.6、4.3、5.8、5.1、8.9个百分点;模型可视化结果显示，改进模型在分散、密集、甘薯苗遮挡场景的识别效果优于其他模型。改进模型部署在NVIDIA Jetson Orin Nano上的田间试验结果表明，mAP50为94.6%，单幅图像平均推理时间为33.26 ms，满足甘薯大田复杂环境移动设备部署和实时性需求，可为后续甘薯大田除草机器人产业化应用提供技术支撑。

Abstract:

Aiming at the problem that weeds in small targets and overlapping occluded areas were difficult to identify under the complex environment of sweet potato field planting, and that the real-time performance was poor due to limited computing power of mobile devices, the YOLO 11n model was improved and a lightweight model (SCSF-YOLO 11n) was proposed for sweet potato seedling and weed identification that balanced accuracy and speed.Firstly, the ShuffleNetV2 lightweight network was adopted as the backbone feature extraction network to reduce the number of parameters and floating-point operations of the model.Secondly, the Context Guided Block was used to replace the bottleneck structure of the C3k2 module in the neck network, integrating the target area and surrounding environment information to reduce the interference of field background.To solve the problem of missing local detailed features caused by overlapping occlusion between sweet potato seedlings and weeds, the separated and enhancement attention module (SEAM) attention mechanism was integrated into the head network to extract and fuse multi-scale features, improving the identification ability of the model in occluded scenes.Finally, Focaler-MPDIoU was adopted as the loss function to accelerate the convergence speed of the model network and reduce its loss value.Experimental results showed that the improved model achieved precision, recall and mAP50 (mean average precision at 0.5 IoU) of 93.6%, 94.3% and 95.4% on the validation set, respectively, with only 1.4×10^6 parameters, 3.2 MB memory occupation, 3.1×10^9 floating-point operations and a processing speed of 112.6 frames per second.Compared with YOLO v5n, YOLOX-tiny, YOLO v7-tiny, YOLO v8n, YOLO v9t, YOLO v10n, YOLO 12n and DINO, the mAP50 was increased by 3.0, 4.1, 3.7, 4.6, 4.3, 5.8, 5.1 and 8.9 percentage points, respectively.Visualization results demonstrated that the improved model outperformed other models in scattered, dense and sweet potato seedling occluded scenes.Field tests deploying the improved model on NVIDIA Jetson Orin Nano showed that the mAP50 was 94.6%, and the average inference time for a single image was 33.26 ms, which met the requirements of mobile device deployment and real-time performance in complex sweet potato field environments.The research result can provide technical support for the subsequent industrial application of sweet potato field weeding robots.