Deep Learning-Based Automated Approach for Determination of Pig Carcass Traits

Abstract

Pig carcass traits are among the most economically significant characteristics and are crucial for genetic selection in breeding and enhancing the economic efficiency. Standardized and automated carcass phenotyping can greatly enhance the measurement efficiency and accuracy, thereby facilitating the selection and breeding of superior pig carcasses. In this study, we utilized phenotypic images and data from 3912 pigs to propose a deep learning-based approach for the automated determination of pig carcass phenotypic traits. Using the YOLOv8 algorithm, our carcass length determination model achieves an average accuracy of 99% on the test set. Additionally, our backfat segmentation model, YOLOV8n-seg, demonstrates robust segmentation performance, with a Mean IoU of 89.10. An analysis of the data distribution comparing manual and model-derived measurements revealed that differences in the carcass straight length are primarily concentrated between −2 cm and 4 cm, while differences in the carcass diagonal length are concentrated between −3 cm and 2 cm. To validate the method, we compared model measurements with manually obtained data, achieving coefficients of determination (R2) of 0.9164 for the carcass straight length, 0.9325 for the carcass diagonal length, and 0.7137 for the backfat thickness, indicating high reliability. Our findings provide valuable insights into automating carcass phenotype determination and grading in pig production.

Keywords

• deep learning
• automated phenotyping of pigs
• carcass length
• backfat thickness