Multi-Scale Feature Fusion and Attention-Enhanced R2U-Net for Dynamic Weight Monitoring of Chicken Carcasses

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Manual collection of broiler chicken weight is inefficient and labor-intensive.
• Accurate weight monitoring is vital for assessing growth and health in poultry.
• Automated systems are needed to improve efficiency in broiler production.

Purpose of the Study:

• To develop a deep learning model for automated broiler carcass weight detection.
• To improve the accuracy and efficiency of weight estimation in yellow-feather broilers.
• To reduce the burden of traditional manual weighing methods.

Main Methods:

• A novel segmentation network (AR2U-AtNet) was developed, incorporating Convolutional Block Attention Module (CBAM) and Selective Kernel Attention (SKAttention).
• A custom dataset of 2709 images from 301 Taihu yellow chickens was utilized.
• The model segments broiler carcasses and uses pixel area for weight prediction via regression.

Main Results:

• The segmentation network achieved an average Intersection over Union (mIoU) of 90.45%, with Dice and F1 scores of 95.18%.
• The regression model for weight prediction obtained an R2 value of 0.9324.
• The proposed method demonstrated high accuracy and speed in determining individual broiler carcass weights.

Conclusions:

• The deep learning model effectively automates broiler carcass weight estimation.
• This approach significantly enhances production efficiency for yellow-feather broilers.
• The findings alleviate the limitations of traditional manual weighing techniques.