Deep Learning and Machine Vision Approaches for Posture Detection of Individual Pigs.

Area of Science:

• Agricultural Science
• Computer Vision
• Animal Welfare

Background:

• Accurate posture detection is crucial for monitoring pig health and welfare.
• Existing machine vision methods often rely on 3D systems or 2D systems under controlled conditions, limiting practical application.

Purpose of the Study:

• To evaluate the efficacy of a 2D imaging system combined with deep learning for detecting pig postures (standing, lying on side, lying on belly) in commercial farm environments.
• To address the limitations of existing posture detection technologies in real-world settings.

Main Methods:

• Proposed three deep learning-based object detection methods: Faster R-CNN, SSD, and R-FCN.
• Integrated feature extraction using Inception V2, ResNet, and Inception ResNet V2 with RGB images.
• Trained and validated models using data from diverse commercial pig farms.

Main Results:

• The R-FCN ResNet101 model demonstrated superior performance in posture detection.
• Achieved high average precision (AP) scores: 0.93 for standing, 0.95 for lying on side, and 0.92 for lying on belly.
• Obtained a mean average precision (mAP) exceeding 0.93, indicating robust detection capabilities.

Conclusions:

• A 2D imaging system coupled with deep learning, specifically the R-FCN ResNet101 method, is effective for detecting pig postures under commercial farm conditions.
• This approach offers a viable, non-invasive solution for continuous monitoring of pig welfare in real-world agricultural settings.