基于改进YOLO v5s的奶山羊面部识别方法

为准确高效地实现无接触式奶山羊个体识别，以圈养环境下奶山羊面部图像为研究对象，提出一种基于改进YOLO v5s的奶山羊个体识别方法。首先，从网络上随机采集350幅羊脸图像构成羊脸面部检测数据集，使用迁移学习思想预训练YOLO v5s模型，使其能够检测羊脸位置。其次，构建包含31头奶山羊3 844幅不同生长期的面部图像数据集，基于预训练的YOLO v5s,在特征提取层中引入SimAM注意力模块，增强模型的学习能力，并在特征融合层引入CARAFE上采样模块以更好地恢复面部细节，提升模型对奶山羊个体面部的识别精度。实验结果表明，改进YOLO v5s模型平均精度均值为97.41%,比Faster R-CNN、SSD、YOLO v4模型分别提高6.33、8.22、15.95个百分点，比YOLO v5s模型高2.21个百分点，改进模型检测速度为56.00 f/s,模型内存占用量为14.45 MB。本文方法能够准确识别具有相似面部特征的奶山羊个体，为智慧养殖中的家畜个体识别提供了一种方法支持。

Face Recognition Method of Dairy Goat Based on Improved YOLO v5s

In order to accurately and efficiently realize the contactless individual identification of dairy goats, a dairy goat individual identification method based on improved YOLO v5s was proposed by taking the facial images of dairy goats in captive environment as the research object. Firstly, totally 350 sheep face images were randomly collected from the network to form a sheep face facial detection dataset, and the YOLO v5s model was pre-trained by using the transfer learning idea to enable it to detect sheep face positions. Secondly, a facial image dataset was constructed, containing 3844 different growth stages of 31 dairy goats, based on pre trained YOLO v5s, SimAM attention module was introduced in the feature extraction layer to enhance the learning ability of the model, and CARAFE was introduced in the feature fusion layer. The sampling module can better restore facial details and improve the recognition accuracy of the model for individual faces of dairy goats. The experimental results showed that the average accuracy of the improved YOLO v5s model was 97.41%, which was 6.33 percentage points, 8.22 percentage points and 15.95 percentage points higher than that of the Faster R-CNN, SSD and YOLO v4 models, respectively, and 2.21 percentage points higher than that of the original YOLO v5s model. The detection speed of the improved model was 56.00f/s, and the model size was 14.45MB. The method proposed can accurately identify dairy goat individuals with similar facial features, which provided a method support for the identification of livestock individuals in smart farming.