Abstract:

Weeding is a crucial method for boosting crop yield， and the precise identification of weed species is imperative for effective weed control. In this study， the SXAU weed dataset was established and the weather data enhancement and online data augmentation were combined to enhance the weed feature extraction capability of Yolov8. This enhancement aimed to improve the model’s practical application effectiveness while concurrently reducing the power consumption of computer hardware. Through the implementation of transfer learning， the Yolov8 model was initially pre-trained on a public weed dataset and subsequently fine-tuned using the SXAU weed dataset to swiftly acquire the relevant parameters needed by the neural network for weed feature recognition. Employing the Yolov8 algorithm in conjunction with the RealSense D435i depth camera， this approach involved utilizing a deep learning algorithm to derive the two-dimensional coordinates of crops and weeds within the image. Subsequently， the depth camera was employed to acquire three-dimensional coordinate information， facilitating spatial positioning of the crops and weeds. Experimental results demonstrate that， following the incorporation of weather data enhancement and fine-tuning， the model achieved an average detection accuracy of 97.43% and F1 value of 94.82%. The average detection time was 13.032 ms， surpassing performance metrics of Yolov7-tiny， Yolov5， Effcientdet， and other models. This research offered valuable insights for studies on lightweight weed identification， as well as the identification and spatial positioning of intelligent weeding robots.

Key words: weed feature extraction, weather data enhancement, transfer learning, Yolov8, lightweight, depth cemera

摘要：

除草是提高作物产量的重要手段，而精准识别杂草种类是精确除草的前提。构建了SXAU杂草数据集，采用天气数据增强与在线数据增强相结合的方法提升Yolov8对杂草特征的提取能力，从而增强模型在实际应用中的效果，同时降低了计算机硬件的功耗。引入迁移学习，在公开杂草数据集上对Yolov8模型进行预训练，使用SXAU杂草数据集微调，以快速获取神经网络对杂草特征所需的相关参数。结合Yolov8算法和RealSense D435i深度相机，通过深度学习算法获取图像中作物与杂草的二维坐标，再利用深度相机得到目标三维坐标信息，实现对作物与杂草的空间定位。结果表明，基于天气数据增强和微调后的模型，平均检测精度和F1值分别达到97.43%和94.82%，平均检测时间为13.032 ms，优于Yolov7-tiny、Yolov5、Effcientdet等模型。研究结果可为轻量级杂草识别研究与智能除草机器人识别定位提供参考。

关键词: 杂草特征提取, 天气数据增强, 迁移学习, Yolov8, 轻量级, 深度相机