A Portable Device for Obtaining Body Condition Score of Dairy Cattle Based on Image Processing and Convolutional Neural Networks

The present work develops an image classifier algorithm to measure the body condition score in Holstein cows. The algorithm aims to reduce the subjectivity that arises when evaluating cattle through visual inspection by specialists. This score measures how thin or overweight are cows in stables, which impacts milk production and the quality of life of the cattle. Although state-of-the-art attempts to solve the subjectivity problem, an efficient and satisfactory method for classification has not yet been found. Moreover, implementations have only considered placing fixed devices in the stables under certain restrictions. Therefore, a portable device with a graphical user interface was designed, and the images were captured and then segmented in a DeepLab3 + convolutional neural network. With this segmented database, the classifier algorithm was trained. For the validation of image segmentation, the Coefficient of Intersection over Union was used, achieving results over 0.9. This finally allowed us to obtain satisfactory results in the calculation of the body condition score.