Interactive Dairy Goat Image Segmentation for Precision Livestock Farming.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Agricultural Technology

Background:

• Deep learning-based semantic and instance segmentation are crucial for intelligent dairy goat farming.
• Current methods like Labelme require extensive pixel-level annotations, proving inefficient and time-consuming.
• High-quality annotations are essential for training accurate segmentation models.

Purpose of the Study:

• To reduce the annotation workload for dairy goat images.
• To improve the segmentation accuracy of deep learning models, particularly on object boundaries and small objects.
• To develop an efficient dairy goat image annotation system.

Main Methods:

• Proposed a novel interactive segmentation model, UA-MHFF-DeepLabv3+, incorporating layer-by-layer multi-head feature fusion (MHFF) and upsampling attention (UA).
• Enhanced the DeepLabv3+ architecture to improve segmentation of object boundaries and small objects.
• Designed and developed a dedicated dairy goat image annotation system named DGAnnotation.

Main Results:

• The UA-MHFF-DeepLabv3+ model achieved state-of-the-art segmentation accuracy on the DGImgs dataset.
• Achieved significantly lower mNoC@85 (1.87) and mNoC@90 (4.11) compared to previous models (3 and 5).
• The DGAnnotation system demonstrated a five-fold increase in annotation speed, annotating a dairy goat instance in just 7.12 seconds.

Conclusions:

• The proposed UA-MHFF-DeepLabv3+ model effectively improves segmentation accuracy in dairy goat farming applications.
• The DGAnnotation system substantially reduces the time and effort required for pixel-level image annotation.
• These advancements facilitate the practical implementation of deep learning in intelligent dairy farming.