Authors

Abstract

Accurate identification of sheep is of significant importance for modern, intensive sheep farming. Traditionally, herders have used conventional methods to identify individual sheep, which are time-consuming, labor-intensive, and prone to considerable errors. In recent years, researchers have developed sheep face recognition models based on deep learning techniques to identify sheep using facial images. However, existing models suffer from insufficient theoretical research and limited recognition accuracy. To address these issues, this study develops a high-accuracy sheep face recognition model named ResNet-SFR. The core innovation of this model is the deepening of the feature extraction network of the original ResNet50, which enhances the model's ability to capture various facial features in sheep images, as well as improving its generalization and stability. Additionally, the Convolutional Block Attention Module (CBAM) attention mechanism is embedded into the original model to further enhance the identification of key features, significantly increasing the accuracy of sheep face recognition. Transfer learning is employed to pre-train the sheep face recognition model, further boosting the accuracy of ResNet-SFR. Experimental results show that on a self-constructed sheep face image dataset, ResNet-SFR achieves a recognition accuracy of 96.6%, demonstrating its superior performance in sheep face recognition tasks. The proposed ResNet-SFR not only offers high recognition accuracy but also exhibits strong applicability, meeting the practical needs of farm identification and showcasing promising application prospects.

Abstract in Chinese

羊只身份准确识别对现代化、集约化养羊业有着重要的应用意义。在过去，牧民们用传统的羊只身份识别方法对个体羊只进行身份识别，然而，传统方法既费时又费力，还存在较大的识别误差。近年来，学者们基于深度学习技术开发了羊脸识别模型，通过羊脸图像识别其对应的身份。然而，目前现有的羊脸识别模型存在理论研究不足和识别精度不足问题。针对上述问题，本研究开发了一组高精度羊脸识别模型，名为ResNet-SFR。该模型的核心创新在于将原ResNet50的特征提取网络进行加深，此举不仅增强了模型捕捉羊脸图像中不同脸部特征的能力，同时也提高了其泛化性与稳定性。此外，在原模型的基础上嵌入了CBAM注意力机制，进一步加强了模型对关键特征的识别，显著提高了羊脸识别的准确度。本研究采用了迁移学习对羊脸识别模型进行预训练，进一步提升了ResNet-SFR的识别精度。试验结果表明，在自制的羊脸图像数据集上，ResNet-SFR的识别精度达到了96.6%，证明了其在应对羊脸识别任务上的优越表现。本研究提出的ResNet-SFR在羊脸识别方面不仅识别精度高，且具有较强的应用性，符合养殖场识别的实际需求，展现了较好的应用前景。