DASNet a dual branch multi level attention sheep counting network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Agricultural Technology

Background:

• Accurate sheep counting is vital for livestock management and grassland ecosystem health.
• Traditional methods are labor-intensive and costly, especially for dense herds.
• Existing computer vision methods struggle with overcounting, missed detections, and extensive annotation needs.

Purpose of the Study:

• To develop a more accurate and efficient automated sheep counting system for grasslands.
• To address challenges in counting sheep from drone imagery, including background interference and varying densities.
• To introduce a novel deep learning model and a diverse, real-world dataset for sheep counting.

Main Methods:

• Collected the Sheep1500 UAV Dataset featuring diverse grassland scenes and sheep densities captured by drones.
• Proposed DASNet, a dual-branch multi-level attention network based on density map regression.
• Integrated a Conv Convolutional Block Attention Layer (CCBL) and a Multi-Level Attention Module (MAM) with residual connections for enhanced feature extraction and fusion.

Main Results:

• DASNet significantly outperformed the baseline MAN network on the Sheep1500 UAV Dataset.
• Achieved a Mean Absolute Error (MAE) of 3.95 and Mean Squared Error (MSE) of 4.87.
• Demonstrated effectiveness in challenging scenarios like dense flocks and background noise.

Conclusions:

• DASNet provides an accurate and computationally efficient solution for automated sheep counting in precision agriculture.
• The proposed model and dataset advance the capabilities of computer vision in ecological and agricultural monitoring.
• DASNet's dual-branch and multi-level attention mechanisms enhance feature representation for improved counting accuracy.