Motion blur aware multiscale adaptive cascade framework for ear tag dropout detection in reserve breeding pigs

Area of Science:

• Agricultural Science
• Computer Vision
• Animal Science

Background:

• Ear tag dropout is a significant issue in precision breeding and health monitoring of active pigs.
• Challenges in detection include motion blur, small tag size, and scale variations.

Purpose of the Study:

• To develop an accurate and efficient method for detecting ear tag dropout in reserve breeding pigs.
• To address the limitations of existing methods in handling motion blur and scale variations.

Main Methods:

• Proposed a motion blur-aware multi-scale framework named Adapt-Cascade.
• Utilized a Weight-Adaptive Attention Module (WAAM) for motion blur feature extraction.
• Employed Density-Aware Dilated Convolution (DA-DC) for small tag detection and Feature-Guided Multi-Scale Region Proposal (FGMS-RP) for multi-scale detection.

Main Results:

• Adapt-Cascade achieved 93.46% accuracy in detecting ear tag dropout.
• The model operates at 19.2 frames per second, enabling real-time application.
• Demonstrated superior performance in handling motion blur, small tag sizes, and scale variations.

Conclusions:

• Adapt-Cascade offers a high-accuracy solution for automated ear tag dropout detection in pigs.
• The developed model supports intelligent pig farm management through improved monitoring.
• This framework enhances the reliability of data for precision breeding and health evaluations.