基于脸部RGB-D图像的牛只个体识别方法

为实现非接触、高精度个体识别，本文提出了一种基于牛只脸部RGB-D信息融合的个体身份识别方法。以108头28~30月龄荷斯坦奶牛作为研究对象，利用Intel RealSense D455深度相机采集2334幅牛脸彩色/深度图像作为原始数据集。首先，采用冗余图像剔除方法和自适应阈值背景分离算法进行图像预处理，经增强共得到8344幅牛脸图像作为数据集；然后，分别选取Inception ResNet v1、Inception ResNet v2和SqueezeNet共3种特征提取网络进行奶牛脸部特征提取研究，通过对比分析，确定FaceNet模型的最优主干特征提取网络；最后，将提取的牛脸图像特征L2正则化，并映射至同一特征空间，训练分类器实现奶牛个体分类。测试结果表明，采用Inception ResNet v2作为FaceNet模型的主干网络特征提取效果最优，在经过背景分离数据预处理的数据集上测试牛脸识别准确率为98.6%，验证率为81.9%，误识率为0.10%。与Inception ResNet v1、SqueezeNet网络相比，准确率分别提高1、2.9个百分点；与未进行背景分离的数据集相比，准确率提高2.3个百分点。

Individual Identification of Cattle Based on RGB-D Images

Individual identification is the foundation for achieving digital management of cattle. In order to achieve non-contact and high-precision individual identification, a dairy cow face recognition method based on RGB-D information fusion was proposed. Totally 108 Holstein cows aged 28 months to 30 months were selected as the research subjects, and 2334 color/depth images of cattle faces were collected by using the Intel RealSense D455 depth camera as the original dataset. Firstly, image preprocessing was carried out by using redundant image elimination and adaptive threshold background separation algorithms. After enhancement, a total of 8344 cattle face images was obtained as the dataset. Then, three feature extraction networks, including Inception ResNet v1, Inception ResNet v2, and SqueezeNet, were selected to extract the facial features of the cattle face. The optimal backbone feature extraction network of the FaceNet model was determined through comparative analysis. Finally, the extracted dairy cow face image features were L2 regularization and mapped to the same feature space. A classifier was trained to achieve individual classification of dairy cows. The test results showed that using Inception ResNet v2 as the backbone feature extraction network of the FaceNet model had the best performance. After testing the cow face recognition accuracy on the preprocessed dataset with background separation, the accuracy reached 98.6%, the verification rate was 81.9%, and the misidentification rate was 0.10%. Compared with that of Inception ResNet v1 and SqueezeNet networks, the accuracy was improved by 1 percentage points and 2.9 percentage points, respectively. Compared with that of the dataset without background separation, the accuracy was improved by 2.3 percentage points. The research result can provide a method for dairy cow face recognition.