Intelligent vineyard blade density measurement method incorporating a lightweight vision transformer

文献类型： 外文期刊

第一作者： Ke, Shan

作者： Ke, Shan;Pan, Hui;Dai, Guowei;Dai, Guowei;Jin, Bowen

作者机构：

关键词： Deep learning; Image processing; Vision transformer; Fusion data augmentation; Leaf density measurement

期刊名称：EGYPTIAN INFORMATICS JOURNAL （ 影响因子：5.2； 五年影响因子：5.0 ）

ISSN： 1110-8665

年卷期： 2024 年 26 卷

页码：

收录情况： SCI

摘要： Under the new demand model of Agriculture 4.0, automated spraying is a very complex task in precision agriculture, which needs to be combined with a computerized vision perception system to distinguish the plant leaf density and execute the spraying operation in real-time accordingly. Aiming at the accurate determination of grape leaf density, an image leaf density determination method based on the lightweight Vision Transformer (ViT) architecture is proposed, which designs a fusion data augmentation method containing a dual augmentation spatial extension and weather data augmentation method, where the former adopts the pixel augmentation and spatial augmentation for the original image processing, and the latter realizes the data augmentation from the empirical point of view adapted to the agricultural operation environment, and fuses the two in order to expand the sample capacity of the grape leaf density image, which then enhances the model's generalization ability and robustness. The lightweight ViT model has self-attention that can automatically and efficiently extract high-frequency local feature representations and use the two-branch structure to mix high-frequency and lowfrequency information to form grapevine-leaf density features in the region of interest. The semantic analysis of the feature extraction layer is parsed using t-SNE and histogram methods, which improves the transparency of the model from the multidimensional with frequency domain distribution space. The experimental results show that the fusion data augmentation method can effectively improve the model recognition accuracy, and the accuracy of comparing the included data augmentation methods is improved by 0.55 % and 3.46 %, respectively. The accuracy of recognizing all four types of grape leaf densities exceeded 94 %, and the MCC reached 90.39 %. In addition, the proposed lightweight ViT improves the accuracy by at least 0.34 % with FLOPs of only 0.6 G compared to the popular MobileViT. The proposed method of this work has high recognition speed and accuracy, which can provide practical technical support for plant protection spraying robots and improve the profitability of growers based on the reduction of pesticide residues.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序