基于注意力机制和边缘感知的田梗提取模型

田埂精确提取是数字化农业管理的重要前提。针对由于遮挡、斑秃等因素干扰，给基于语义分割方法提取田埂带来困难问题，提出一种基于注意力机制和边缘感知模块的U-Net网络实现田埂提取。首先，将多信息注意力引入U型分割网络的下采样中，增强相邻层之间的上下文信息，提升对田埂区域语义特征的表示能力。其次，将边缘感知分割模块应用至U-Net解码部分的每一层，在不同语义特征层提取田埂边缘信息，提高田埂区域语义分割精度。最后，联合边缘感知损失与语义分割损失构建联合损失函数，用于整体网络优化。通过对安徽省淮北市濉溪县小麦基地采集的无人机麦田数据集进行训练和模型验证，实验结果表明，本文模型语义分割像素准确率高达95.57%,平均交并比达到77.48%。

Ridge Extraction Model Based on Attention Mechanism and Edge Perception

Accurate extraction of ridges is an important prerequisite for digital agricultural management. However, due to the interference of factors such as occlusion and alopecia areata, it brings challenges for the semantic segmentation method to extract the ridge area. A U-Net segmentation network model was proposed based on a multi-information attention mechanism and an edge-aware module. Firstly, multi-information attention was introduced into the down-sampling of the U-shaped network to enhance the context information between adjacent layers and improve the representation ability of the semantic features of the ridge area. Secondly, the edge-aware segmentation module was applied to each layer of the U-Net decoding part, and the ridge edge information was extracted in different semantic feature layers to improve the semantic segmentation accuracy of the ridge region. Finally, the joint edge-aware loss and semantic segmentation loss were used to construct a joint loss function for overall network optimization. The training and model validation were carried out with the UAV wheat field data set collected by the wheat experimental base in Suixi County, Huaibei City, Anhui Province. The experimental results showed that the pixel accuracy of semantic segmentation of crop plants in different datasets was as high as 95.57%, and the average intersection ratio was 77.48%.