野生荒漠草原的特点是栖息地多样,植物分布不均,植物类之间的相似性,和植物阴影的存在。然而,现有的检测荒漠草原植物物种的模型精度低,需要大量的参数,并招致高昂的计算成本,使它们不适合在这些环境中的工厂识别场景中部署。为了应对这些挑战,本文提出了一种轻量级、快速的植物物种检测系统,称为YOLOv8s-KDT,为复杂的沙漠草原环境量身定制。首先,该模型引入了一种动态卷积KernelWarehouse方法,以降低卷积内核的维数并增加其数量,从而在参数效率和表示能力之间实现更好的平衡。其次,该模型将三元组注意力纳入其特征提取网络,有效地捕捉信道与空间位置的关系,增强模型的特征提取能力。最后,动态探测头的引入解决了与目标探测头和注意力不均匀有关的问题,从而改进目标检测头的表示,同时降低计算成本。实验结果表明,升级后的YOLOv8s-KDT模型能够快速有效地识别荒漠草地植物。与原始模型相比,FLOP下降50.8%,精度提高了4.5%,mAP增加了5.6%。目前,将YOLOv8s-KDT模型部署在宁夏荒漠草原移动植物识别APP和定点生态信息观测平台中。它有助于调查整个宁夏地区的荒漠草原植被分布以及长期观察和跟踪特定地区的植物生态信息,比如大水坑,黄集田,和宁夏的红寺步。
Wild desert grasslands are characterized by diverse habitats, uneven plant distribution, similarities among plant class, and the presence of plant shadows. However, the existing models for detecting plant species in desert grasslands exhibit low precision, require a large number of parameters, and incur high computational cost, rendering them unsuitable for deployment in plant recognition scenarios within these environments. To address these challenges, this paper proposes a lightweight and fast plant species detection system, termed YOLOv8s-KDT, tailored for complex desert grassland environments. Firstly, the model introduces a dynamic convolutional KernelWarehouse method to reduce the dimensionality of convolutional kernels and increase their number, thus achieving a better balance between parameter efficiency and representation ability. Secondly, the model incorporates triplet attention into its feature extraction network, effectively capturing the relationship between channel and spatial position and enhancing the model\'s feature extraction capabilities. Finally, the introduction of a dynamic detection head tackles the issue related to target detection head and attention non-uniformity, thus improving the representation of the target detection head while reducing computational cost. The experimental results demonstrate that the upgraded YOLOv8s-KDT model can rapidly and effectively identify desert grassland plants. Compared to the original model, FLOPs decreased by 50.8%, accuracy improved by 4.5%, and mAP increased by 5.6%. Currently, the YOLOv8s-KDT model is deployed in the mobile plant identification APP of Ningxia desert grassland and the fixed-point ecological information observation platform. It facilitates the investigation of desert grassland vegetation distribution across the entire Ningxia region as well as long-term observation and tracking of plant ecological information in specific areas, such as Dashuikeng, Huangji Field, and Hongsibu in Ningxia.