【摘要】 针对传统检测方法在牛脸检测应用方面存在的检测设备易损、检测结果不理想等问题,根据大数据、多差异性原则,使用手机和相机在某奶牛养殖场采集奶牛数据,构建了一个超过10 000张不同条件下（如遮挡、模糊、光照变化等）的奶牛数据集.在此基础上使用目前有代表性的基于深度网络模型的目标检测方法（如SSD,Faster R-CNN和R-FCN等）对该数据集进行试验对比分析.结果表明:Faster R-CNN模型综合检测精度最高,可达0.990,但其检测速度相对较慢,为11 F·s^-1;SSD模型的检测速度最快,为47 F·s^-1,但其检测精度与Faster R-CNN相比略低,约为0.945.更多还原

【Abstract】 To solve the problems of currently traditional detection methods for cow face detection with poor detection effect and easy damage of detection equipment, according to the principles of big data and diversity, the mobile phones and cameras were used to build the dataset with more than 10 000 cows under different conditions of appearance variation, occlusion and illumination change. Using the dataset, the object detection methods based on deep network models of SSD, Faster R-CNN and R-FCN were improved and compared on the detection performance. The results show that the improved Faster R-CNN can achieve the detection accuracy of 0.990 with detection speed of 11 F·s^-1. The detection speed of the improved SSD is 47 F·s^-1, and the detection accuracy is 0.945, which is slightly lower than that of Faster R-CNN.更多还原