Cattle Target Segmentation Method in Multi-Scenes Using Improved DeepLabV3+ Method.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Animal Science

Background:

• Accurate animal detection and positioning are crucial for understanding animal behavior and advancing smart animal farming.
• Complex breeding environments pose significant challenges to existing semantic segmentation models, leading to poor target segmentation and weak generalization.

Purpose of the Study:

• To develop a more effective semantic segmentation model for complex animal farming scenes.
• To improve target segmentation accuracy and model generalization capabilities.

Main Methods:

• An improved DeepLabV3+ network (Imp-DeepLabV3+) was proposed, replacing the backbone with MobileNetV2 for enhanced feature extraction.
• A layer-by-layer feature fusion method was implemented in the Decoder stage for multi-scale integration of semantic and high-resolution features.
• The Squeeze-and-Excitation (SENet) module was incorporated to improve feature fusion and segmentation precision.

Main Results:

• The Imp-DeepLabV3+ model achieved high performance metrics: 99.4% pixel accuracy (PA), 98.1% mean pixel accuracy (MPA), and 96.8% mean intersection over union (MIoU).
• The improved model demonstrated significantly enhanced segmentation performance compared to the original DeepLabV3+.
• Imp-DeepLabV3+ outperformed other common semantic segmentation models like FCNs, LR-ASPP, and U-Net.

Conclusions:

• The proposed Imp-DeepLabV3+ model offers superior performance for animal segmentation in challenging environments.
• This advancement is highly applicable to scene segmentation tasks, individual information analysis, and the development of intelligent animal farming systems.