Abstract:

Automated image-based phenotyping has become more feasible due to the improvement in deep learning methods aiming at identifying regions of interests within an image. To investigate the possibility of using standard deep learning tools to predict the conformation score out of an image, a dataset of 8,950 images obtained from 3,561 different cows was used. Each animal was also scored by a classifier. The conformation traits analysed included: udder depth, rear leg set and rump angle. These showed that automated extraction of relevant regions within an image are possible and the defined regions can be used to predict a classification score for an animal. Image-based (deep learning) heritability scores were 0.35, 0.20 and 0.25 for udder depth, rear leg set and rump angle, respectively. Genetic correlations between both methods were highly positive for all traits.