Cow tail detection method for body condition score using Faster R-CNN

摘要

Body condition score (BCS) is an important parameter to measure cow energy reserve for feeding management. Currently, measuring BCS mainly relies on veterinary experts or skilled scorers by observing and touching animals, which is time consuming and costly, especially on large-scale farms. In this study, a method of image processing and deep learning is employed to estimate cow BCS. Firstly, the network cameras were used to capture the back view images of the cows, resulting in 5470 images that constituted the sample data set, in which the key body parts (tail, pins and rump) of the cow were labeled manually. Secondly, Faster Region-Convolutional Neural Networks (Faster R-CNN) method was used to position and classify the cow tail images which were correlated to the value of BCS. Compared with other deep learning algorithms like the Single Shot multibox Detector (SSD), Faster R-CNN had a slightly higher accuracy. Specifically, the detection accuracy for cow tails was 84%, and the BCS classification had an average accuracy 70%. The low BCS classification rate was mainly due to the insufficient image data in this study for cows with certain ranges of BCS.