Beef Carcass Grading using Deep Convolutional Networks

Abstract

Beef carcass grading is an invaluable tool to ensure meat quality. In most of the Brazilian abattoirs, carcasses are graded through visual analysis by trained graders. In order to automate this process, we evaluate seven image deep learning models. For this purpose, a new dataset was created containing images of 670 bovine half-carcasses taken during regular operation in an abattoir. The images were graded by three professionals. All three experts agreed in only 9.9% of the cases, and two out of three graders agreed in 58.82%. The graders disagreed on 31.28% of the images. These results indicate the complexity of the problem. Nonetheless, an overall accuracy of 53% was achieved using convolutional neural networks, which is close to human performance, when the agreement between the graders is considered. Furthermore, an accuracy of around 91% can be achieved if the cases of disagreement are disregarded.