G. Azzaro ,* M. Caccamo ,*1 J. D. Ferguson ,† S. Battiato ,‡ G. M. Farinella ,‡ G. C. Guarnera ,‡ G. Puglisi ,‡
R. Petriglieri ,* and G. Licitra *§

* – Consorzio Ricerca Filiera Lattiero-Casearia (CoRFiLaC), Regione Siciliana, 97100 Ragusa, Italy
†-  University of Pennsylvania, School of Veterinary Medicine, Kennett Square 19348
‡-  Image Processing Laboratory, University of Catania, 95125 Catania, Italy
§-  Dipartimento di Scienze Agronomiche, Agrochimiche e di Produzioni Animali (D.A.C.P.A.), Agriculture Faculty, University of Catania, 95123 Catania, Italy

Body condition score (BCS) is considered an important tool for management of dairy cattle. The feasibility of estimating the BCS from digital images has been demonstrated in recent work. Regression machines have been successfully employed for automatic BCS estimation, taking into account information of the overall shape or information extracted on anatomical points of the shape. Despite the progress in this research area, such studies have not addressed the problem of modeling the shape of cows to build a robust descriptor for automatic BCS estimation. Moreover, a benchmark data set of images meant as a point of reference for quantitative evaluation and comparison of different automatic estimation methods for BCS is lacking. The main objective of this study was to develop a technique that was able to describe the body shape of cows in a reconstructive way. Images, used to build a benchmark data set for developing an automatic system for BCS, were taken using a camera placed above an exit gate from the milking robot. The camera was positioned at 3 m from the ground and in such a position to capture images of the rear, dorsal pelvic, and loin area of cows. The BCS of each cow was estimated on site by 2 technicians and associated to the cow images. The benchmark data set contained 286 images with associated BCS, anatomical points, and shapes. It was used for quantitative evaluation. A set of example cow body shapes was created. Linear and polynomial kernel principal component analysis was used to reconstruct shapes of
cows using a linear combination of basic shapes constructed from the example database. In this manner, a cow’s body shape was described by considering her variability from the average shape. The method produced a compact description of the shape to be used for automatic estimation of BCS. Model validation showed that the polynomial model proposed in this study performs better (error = 0.31) than other state-of-the-art methods in estimating BCS even at the extreme values of BCS scale.

Keywords: body condition score, digital imaging, body shape