基于MSRCP与改进YOLO v4的躺卧奶牛个体识别方法

Individual Identification Method of Lying Cows Based on MSRCP and Improved YOLO v4 Model

奶牛的躺卧率可以反映奶牛的舒适度和健康情况,躺卧奶牛的个体识别是自动监测奶牛躺卧率的基础。本文提出了一种基于改进YOLO v4模型识别非限制环境下躺卧奶牛个体的方法。为实现对躺卧奶牛全天的准确个体识别,首先对18:00—07:00的图像采用MSRCP(Multi-scale retinex with chromaticity preservation)算法进行图像增强,改善低光照环境下的图像质量。其次,在YOLO v4模型的主干网络中融入RFB-s结构,改善模型对奶牛身体花纹变化的鲁棒性。最后,为提高模型对身体花纹相似奶牛的识别准确率,改进了原模型的非极大抑制(Non-maximum suppression,NMS)算法。利用72头奶牛的图像数据集进行了奶牛个体识别实验。结果表明,相对于YOLO v4模型,在未降低处理速度的前提下,本文改进YOLO v4模型的精准率、召回率、mAP、F1值分别提高4.66、3.07、4.20、3.83个百分点。本文研究结果为奶牛精细化养殖中奶牛健康监测提供了一种有效的技术支持。

The lying rate of dairy cows can reflect the comfort and health of dairy cows.The individual identification of lying cows is the basis of automatic monitoring of lying rate.A method based on the improved YOLO v4 model to identify individual lying cows in an unconstrained barn environment was proposed.Firstly,in order to realize accurate individual identification of lying cows throughout the day,MSRCP algorithm was used to enhance the images from 18:00 to 07:00 the next day,which improved the image quality in low light environment.Secondly,the RFB-s structure was integrated into the backbone network of YOLO v4 model to increase the robustness of the model to the changes of cow body patterns.Finally,in order to improve the identification accuracy rate of cows with similar patterns,the non-maximum suppression(NMS)algorithm of YOLO v4 model was improved.The experiment of cow individual identification was carried out on the image data set of 72 cows.The results showed that the precision,recall,mAP,and F1 values of the improved YOLO v4 were 97.84%,93.68%,96.87%,and 95.71%,respectively.The improved YOLO v4 model was compared with the YOLO v4 model,the precision,recall,mAP and F1 values of the improved YOLO v4 were increased by 4.66 percentage points,3.07 percentage points,4.20 percentage points and 3.83 percentage points,respectively,without reducing the processing speed.The mAP of the improved YOLO v4 was higher than that of YOLO v5,SSD,CenterNet and Faster R-CNN by 8.52 percentage points,15.22 percentage points,12.18 percentage points and 1.55 percentage points,respectively.The method can provide an effective technical support for the health monitoring of dairy cows in precision dairy farming.