基于深度学习的牛脸识别系统设计

Design of cattle face recognition system based on deep learning

为了提高畜牧保险理赔行业中牛只身份识别准确率,试验建立了基于计算机视觉提取牛脸特征的牛只身份识别系统,即使用SOLOv2实例分割模型提取牛脸前景图像,结合FaceNet提取牛脸特征;基于Tornado Web和TF Serving框架进行实例分割模型和牛脸特征提取模型的部署,完成牛脸身份识别系统的搭建,并制作了手机APP,最后对添加SOLOv2牛脸身份识别模型的准确率和验证率进行了验证。结果表明:添加了SOLOv2模型的牛脸身份识别准确率达到了98.063%,验证率达到了92.451%,与未添加SOLOv2模型相比,准确率提高了0.270百分点,验证率提高了6.275百分点。说明添加SOLOv2实例分割模型能提升牛脸身份识别的准确率和验证率。

In order to improve the accuracy of cattle identity recognition in the claims of livestock insurance industry,a cattle identity recognition system based on the extraction of cattle face features by computer vision was established,which extracted the foreground image of cattle face using the SOLOv2 instance segmentation model and the features of cattle face by FaceNet.Based on Tornado Web and TF Serving framework,the case segmentation model and cow face feature extraction model were deployed to complete the establishment of a cow face identity recognition system,and a mobile APP was developed,and the accuracy and velidation rate of adding the SOLOv2 cow face identity recognition model was verified.The results showed that the accuracy rate of cow face identification with SOLOv2 model reached 98.063%,and the validation rate was 92.451%.Compared with the model without SOLOv2,its accuracy rate was improved by 0.270 percentage points and its verification rate was improved by 6.275%.The results indicated that adding SOLOv2 instance segmentation model could improve the accuracy and validation rate of cow face identity recognition.