Utilização da visão computacional para predição do rendimento de cortes da carcaça bovina

Abstract

The objective of this study was to develop prediction equations for bovine carcass cuts using carcass images. The study used twenty Nelore male animals that were slaughtered at an average age of 18 months. The average live weight was 578.80 kg ±103.20 kg, and the average hot carcass weight was 291.98 kg. After cooling, the carcasses were sawed between the 5th and 6th ribs. Ten images were taken per animal using a Xiaomi smartphone attached to a 'ring light' support with standardized light intensity. The carcasses were then separated into forequarters and hindquarters, weighed, deboned, and used to calculate correlations with the tissue areas traced in the images. The Labelme software was used to interpret the images and identify the areas of muscle, bone, fat, and cartilage tissue in each image. The set of images was divided into three groups (training, test, and validation) using cross-validation to evaluate the generalization capacity of the machine learning models. The predictive capacity of the models was assessed using metrics such as Intersection over Union (IoU), recall, precision, and Fscore. The predictive capacity of the models was assessed using the Intersection over Union (IoU), recall, precision, and Fscore metrics. The precision and accuracy values, respectively, in descending order, were obtained for bottom (97.98; 98.0), meat (90.87; 92.71), fat (88.87; 73.50), bone (85.29; 80.52), and cartilage (81.46; 79.15). The lower accuracy values for cartilage, bone, and fat may be due to the similarity in color of the three tissues, which are all light in tone and tend towards white. This similarity may have made it more difficult for the algorithms to differentiate between them. The weights of the forequarters (0.49) and hindquarters (0.25) were positively correlated with the areas of muscle tissue and bone, respectively. The weight of the hind and forequarter trimmings showed a negative correlation with the muscle area (-0.48 and -0.69, respectively). The computer vision algorithms effectively differentiated carcass tissues from images. The correlation between tissue areas and the weight of cuts and trimmings was moderate. This relationship should be further explored in future studies to better understand the potential of computer vision algorithms in predicting the yield of beef carcass cuts.

Keyword: Beef cattle, Carcass, Images, Artificial Intelligence.