Aiming to achieve an effective identification of dairy cows in a noncontact and highprecision environment of farming, a method to identify dairy cows based on the improved YOLO v3 deep convolutional neural network was proposed. According to this method, multiple cameras were installed above the passageway between the doors of the milking room. The back of cows was videotaped automatically and regularly, after which the image of the cows back was captured by applying video frame decomposition technology. Upon the removal of images noise with bilateral filters and the enhancement of brightness and contrast with the pixel linear transformation method, the individual dairy cows were serial numbered manually. For the cows to be better identified in complex environments, the YOLO v3 recognition model that features optimized anchor boxes and improved network structure was constructed by making reference to the Gaussian YOLO v3 algorithm. From totally 36790 images showing the back of 89 cows, 22074 were randomly selected as the training set, while the remaining ones were classified into either the validation set or the test set. The results showed that the accuracy of the improved YOLO v3 was 9591%, the recall rate was 95.32%, the mAP was 95.16%, the IoU was 85.28%, the actual frame rate of detection was 32f/s, and the accuracy rate of identification was 0.94 percentage points higher compared with that of the YOLO v3 and 1.90 percentage points higher than that of Faster R-CNN. Moreover, the detection speed was eight times faster than that of Faster R-CNN, while the F1 value of dairy cows with pure black back was 2.75 percentage points higher compared with that of the original algorithm. The method showed such advantages as low cost and excellent performance, which were not only conducive to the realtime identification of dairy cows in complex farm environments, but also to the extended application of this method to the identification of other largesized animals. 