基于VGG-ST模型的奶牛粪便形态分类方法研究

快速准确识别奶牛粪便形态，对于奶牛肠胃健康监测与精细管理具有重要意义。针对目前奶牛粪便识别人工依赖强、识别难度大等问题，提出了一种基于VGG-ST（VGG-Swin Transformer）模型的奶牛稀便、软便、硬便及正常粪便图像识别与分类方法。首先，以泌乳期荷斯坦奶牛粪便为研究对象，采集上述4种不同形态的粪便图像共879幅，利用翻转、旋转等图像增强操作扩充至5580幅作为本研究数据集；然后，分别选取Swin Transformer、AlexNet、ResNet-34、ShuffleNet和MobileNet 5种典型深度学习图像分类模型进行奶牛粪便形态分类研究，通过对比分析，确定Swin Transformer为最优基础分类模型；最后，融合VGG模型与Swin Transformer模型，构建了VGG-ST模型，其中，VGG模型获取奶牛粪便局部特征，同时Swin Transformer模型提取全局自注意力特征，特征融合后实〖JP3〗现奶牛粪便图像分类。实验结果表明，Swin Transformer模型在测试集中分类准确率达859%，与ShuffleNet、ResNet-34、MobileNet、AlexNet模型相比分别提高1.8、4.0、12.8、23.4个百分点；VGG-ST模型分类准确率达89.5%，与原Swin Transformer模型相比提高3.6个百分点。该研究可为奶牛粪便形态自动筛查机器人研发提供方法参考。

Cow Manure Classification Method Based on VGG-ST Model

Accurate and rapid identification of cow manure morphology is of great significance for monitoring and precise management of cow gastrointestinal health. In response to the problems of strong artificial dependence and difficulty in identification in current cow manure recognition methods, a method for identifying cow thin, loose, hard, and normal manure was proposed based on the VGG-ST (VGG-Swin Transformer) model. Firstly, a total of 879 images of the four different forms of manures was collected from lactating Holstein cows and augmented to 5580 images using operations such as flipping and rotation as the dataset. Then, five typical deep learning image classification models, namely Swin Transformer, AlexNet, ResNet-34, ShuffleNet and MobileNet, were selected for cow manure classification research. Through comparative analysis, Swin Transformer was determined to be the optimal base classification model. Finally, the VGG-ST model combined the VGG model with the Swin Transformer model. The VGG model was utilized to capture local features of cow manure, while the Swin Transformer model extracted global self-attention features. After feature concatenation, the cow manure images were classified. The experimental results showed that the Swin Transformer model achieved a classification accuracy of 85.9% on the testing set, which was 1.8 percentage points, 4.0 percentage points, 12.8 percentage points, and 23.4 percentage points higher than that of ShuffleNet, ResNet-34, MobileNet, and AlexNet, respectively. The classification accuracy of the VGG-ST model was 89.5%, which was 3.6 percentage points higher than that of the original Swin Transformer model. The research result provided a method reference for the development of automatic inspection robots for cow manure morphology.