ABSTRACT
A Computer Vision System prototype for grading pork carcasses was developed at the Lacombe Research System. The system consists of two components: ultrasound imaging to scan a cross-section of the loin muscle and video imaging to capture two-dimensional (2D) and three-dimensional (3D) images of the carcass. For each of the 241 carcasses (114 barrows and 127 gilts), salable meat yield was determined from a full cutout. Linear, two- and three-dimensional, angular and curvature measurements and carcass volume were derived from each image. Muscle area and fat thickness (7 cm off the mid-line) measured by ultrasound at the next to last rib site, together with 2D and 3D measurements provided the most accurate model for estimating salable meat yield (R(2)=0.82 and RSD=1.68). Models incorporating fat thickness and muscle depth measured at the Canadian grading site (3/4 last rib, 7 cm off the mid-line) with the Destron PG-100 probe, had the lowest R(2) and highest residual standard deviation (RSD) values (R(2)=0.66 and RSD=2.15). Cross-validation demonstrated the reliability and stability of the models; hence conferring them good industry applicability. The Lacombe Computer Vision System prototype appears to offer a marked improvement over probes currently used by the Canadian pork industry.
