.png){kind=icon}
.png){kind=icon}
Return
WeedNet-X: A lightweight field weed detection algorithm
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2025.112441.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurately distinguishing field weeds from crops and locating weed positions are critical prerequisites in automated weed control operations. However, weed detection and localization in unstructured field environments with complex lighting remains challenging. Firstly, data-driven deep learning based algorithms usually have a high dependence on a large number of training samples, and there are huge differences between field weeds and crops in different regions, growth cycles, and types. In addition, the conflict between hardware performance and computation cost makes it difficult for existing weed detection algorithms to maintain both detection accuracy and speed on low-performance platforms. All these problems increase the difficulty of detection. To solve the above problems, we first construct a medium-to-large weed dataset using an open-source agricultural image dataset and collect field data. Subsequently, we have proposed a lightweight weed detection algorithm using the ShuffleNetv2 network as the backbone network, with a multi-scale pyramid network, and the overall network algorithm is named WeedNet-X. The number of model parameters and the computational volume of the algorithm are only 0.57 million and 0.48 Giga floating point operations (GFLOPs), respectively. On the two constructed datasets, the mean Average Precision (mAP) of the algorithm can reach 86.31 % and 80.98 %, respectively, which are improved by 0.61 % and 3.10 % compared to the baseline model. Finally, the hardware and software systems for weed detection verify the excellence of the proposed algorithm in terms of practical performance.
Journal
IF:
8
Papers:
5.4K
Citations:
3.5W