Deep learning-based temperature detection for group-raised pigs via thermal images

Body temperature detection of pigs is vital for efficient pig breeding. In this study, a novel approach for individual pig temperature detection in a group-raised environment is proposed. The methodology includes three approaches: (1) a Porcine-ESRGAN algorithm based on generative adversarial networks (GANs) was developed to reconstruct super-resolution of thermal infrared images, effectively addressing low-resolution limitations, (2) a novel temperature extraction method incorporating instantaneous recognition of postures was established to overcome challenges in measuring pigs in the lying position, and (3) an enhanced YOLOv7 architecture (termed ITG-YOLOv7) was implemented to improve the accuracy of thermal region detection. A performance evaluation demonstrated the following significant improvements achieved by the proposed methodology: Porcine-ESRGAN increased precision ( P ), recall ( R ), and mean average precision ( mAP ) for lying-position pigs by 3.1%, 11.5%, and 6.9%, respectively. Furthermore, the posture-based method attained a higher efficiency of 29.69% in the extraction of key-region temperature. ITG-YOLOv7 attained 93.2% P , 95.4% R , and 97.5% mAP when integrated with Porcine-ESRGAN. The mean absolute error ( MAE ) of the measured temperature was 0.09°C for upright pigs and 0.23°C for lying pigs. The results validate the framework’s efficacy for automated, highly precise monitoring of group-housed pigs’ body temperatures.