Instance segmentation of beef carcass features with deep learning

The meat industry is increasingly seeking automation to address challenges such as labour shortages, worker welfare in harsh abattoir environments, strict hygiene requirements for non-contact processing, and the need to reduce waste. However, unlike pork and poultry sectors, automation in beef processing remains limited due to the larger carcass size and greater biological variability. Artificial intelligence (AI), particularly deep learning integrated with traditional computer vision, offers promising capabilities for feature detection in complex and variable objects. This study examines the feasibility of applying deep learning to beef carcass segmentation. Pretrained models from two state-of-the-art frameworks (i.e. You Only Look Once (YOLO) and Detectron2) were retrained, validated, and tested on a diverse dataset of 118 carcasses from a commercial meat plant. Images were captured using two distinct acquisition systems, and models were trained to identify both large anatomical structures (e.g., carcass, aitchbone, spine, ribcage) and smaller features (e.g., individual ribs). Both models achieved higher mean average precision (mAP) for large structures than for smaller ones. The YOLOv8 model achieved mAP50 values (Intersection over Union (IoU)=0.5) between 94.1% and 97.0% for large structures and 78.0% to 92.4% for ribs. Visual assessment showed that smaller features were often missed in background carcasses, leading to lower mAP values, though performance improved under controlled lighting. YOLOv8 outperformed Detectron2 in segmenting rib elements, yielding sharper boundaries. Overall, the findings demonstrate the potential of pretrained deep learning models for beef carcass segmentation, with further improvements expected through hyperparameter tuning, multimodal data integration, and expanded training datasets.