Research on Instance Segmentation Algorithm for Caged Chickens in Infrared Images Based on Improved Mask R-CNN

Area of Science:

• Agricultural Science
• Computer Vision
• Animal Science

Background:

• Infrared imaging offers insights into chicken health.
• Accurate detection and segmentation of individual chickens are crucial for large-scale farming.
• Obstacles and chicken clustering in infrared images pose segmentation challenges.

Purpose of the Study:

• To develop an advanced instance segmentation algorithm for detecting and segmenting caged chickens in infrared images.
• To improve feature extraction and segmentation accuracy in challenging high-density environments.
• To enhance automated health monitoring and intelligent breeding management.

Main Methods:

• A Mask R-CNN-based instance segmentation algorithm was proposed.
• The backbone network was enhanced with CBAM (Convolutional Block Attention Module).
• The algorithm was combined with the AC-FPN (Attentive Cross-scale Feature Pyramid Network) architecture.

Main Results:

• The enhanced model achieved 78.66% AP and 85.80% AR10 for object detection (COCO metrics).
• Segmentation tasks yielded 73.94% AP and 80.42% AR10, improving by 32.91% and 17.78% over the original model.
• The 'Chicken-many' category reached 98.51% segmentation accuracy, with other categories exceeding 93%.

Conclusions:

• The proposed instance segmentation method effectively recognizes and segments caged chickens in challenging infrared imaging conditions.
• This advancement supports enhanced production efficiency and intelligent breeding management in high-density poultry farming.
• The improved accuracy in detecting and segmenting chickens contributes to better health monitoring systems.