【目的】为解决因土壤图像纹理复杂、没有结构性特征导致的传统卷积神经网络模型难以提取其中关键性特征、识别准确率低的问题，提出了一种大核注意力(large kernel attention，LKA)机制模块与ResNet模型融合的土壤种类识别网络模型VAR(visual attention ResNet)，以解决土壤种类样本不平衡和难分类样本造成的模型泛化能力弱的问题。【方法】以ResNet为主干网络，在主干中引入LKA机制，减少其中的残差块，构建土壤种类识别网络VAR，并改进网络的焦点损失函数(Focal Loss)。【结果】1) 与传统模型ResNet18、ResNet34、VGG、GooleNet、VAN等相比，VAR模型在特定模型参数下对紫色土土壤图像数据集中土壤种类的识别精度更高；2) 用3种不同大小VAR模型之一的VAR_small与以ResNet18为主干并嵌入传统注意力机制SE、CBAM、ECA和SK的网络进行对比，实验结果显示LKA机制在土壤识别方面更加优秀；3) 改进的Focal Loss可让VAR更能注意到难分类的土壤图像样本。【结论】将LKA机制模块与ResNet模型融合的土壤种类识别网络模型VAR增强了网络提取土壤图像中关键性结构特征能力，同时还减少了网络参数，能更加有效地识别土壤种类。
［Purposes］Because there are complex textures and no structural characteristic in soil image, it is difficult to extract key features from a soil image with traditional convolutional neural network. A new classification network named visual attention ResNet (VAR), which combined ResNet with large kernel attention (LKA) module, is proposed to solve the low generalization ability that is caused by imbalance datasets and hardly classified samples and it involves VAR_tiny, VAR_small and VAR_base. ［Methods］To construct VAR network for soil classification, the LKA is embedded into ResNet as a part of the backbone and the residual modules is partially removed from ResNet, and an improved focal loss function (Focal Loss) is introduced to VAR. ［Results］1) Compared with the traditional models ResNet18, ResNet34, VGG, GooleNet, VAN, etc., VAR model has a higher identification accuracy for images data set of purple soil species under specific model parameters. 2) VAR_small, one of the three VAR models with different sizes, was compared with ResNet18 networks embedded with traditional attention mechanisms such as SE, CBAM, ECA and SK. The experimental results showed that LKA mechanism was better in soil recognition. 3) The improved Focal Loss enables VAR to pay more attention to soil image samples that are difficult to classify. ［Conclusions］The soil species identification network VAR, which integrates the LKA module and ResNet, enhances the ability of the network to extract the key structural features from soil images, makes it identify soil species more effectively. And it also reduces the network parameters.