SheepInst: A High-Performance Instance Segmentation of Sheep Images Based on Deep Learning.

Area of Science:

• Computer Vision
• Precision Livestock Farming
• Animal Science

Background:

• Sheep instance segmentation is vital for precision livestock farming but challenging due to sheep congregations and irregular shapes.
• Existing methods struggle with accurate individual sheep identification, behavior recognition, and weight estimation in crowded conditions.

Purpose of the Study:

• To develop an advanced sheep instance segmentation method (SheepInst) for improved accuracy in locating and contouring individual sheep, especially in overlapping scenarios.
• To enhance the capabilities of computer vision in sheep farming for better animal management and welfare.

Main Methods:

• Proposed a two-stage sheep instance segmentation model, SheepInst, built upon the Mask R-CNN framework and RefineMask.
• Utilized an improved ConvNeXt-E backbone for feature extraction and a modified Dynamic R-CNN for precise localization of overlapping sheep.
• Integrated spatial attention modules into the RefineMask segmentation network to accurately delineate irregular sheep contours.

Main Results:

• SheepInst achieved high performance metrics on the test set: 89.1% box AP, 91.3% mask AP, and 79.5% boundary AP.
• Demonstrated superior accuracy in segmenting individual sheep, particularly in cases of significant overlap.
• Extensive experiments confirmed the suitability and excellent performance of SheepInst for sheep instance segmentation.

Conclusions:

• SheepInst offers a significant advancement in sheep instance segmentation technology.
• The proposed method effectively addresses the challenges posed by sheep congregations and irregular contours.
• SheepInst is well-suited for integration into precision livestock farming systems, enhancing data acquisition for animal monitoring.