64 / 2025-04-15 11:46:46

Pose Feature Extraction Network(PFEN)-based model to estimate the weight of finishing pig in a free posture by using RGB-D images

pig weight estimation,convnext,pig posture,feature fusion,Coordinate attention

3. 智慧牧场与动物福利 > （2）智慧养殖的大数据、机器学习和物联网应用

Body weight, a crucial parameter in the finishing pig production, can be continuously monitored by using the computer vision-based weight estimation method. However, most of the image data used in existing studies were collected under the feeding conditions of pigs with standing posture, rarely considering the influence of other postures. Actually, the diverse postures are common in actual production environments. In this study, a Pose Feature Extraction Network(PFEN)-based model by using RGB-D images was proposed to estimate the body weight of finishing pig in a free posture. The PFEN introduces a novel structure designed to extract the posture features and integrate them into the weight estimation process. A comprehensive image dataset consisting of 134,810 RGB-D image pairs from 147 different pigs in a real production environment was built to validate the model, which includes three typical pig postures: standing, abdominal lying, and lateral lying. The ablation experimental results show that the mean absolute error is reduced by up to 43.7% with the incorporation of the PFEN structure, thereby confirming the significant impact of posture information on the accurate body weight estimation in a free posture. Based on the proposed PFEN structure, the high-and lowdi-mensional feature fusion method and the coordinate attention module were added to develop the PFEN-based model for pig weight estimation, further reducing the mean absolute error by 9.6%. The optimal pig weight estimation results under the image dataset built in this study achieve a mean absolute error of 5.147 and an R2 of 0.800.