Organ Segmentation in Poultry Viscera Using RGB-D.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Manual inspection of eviscerated poultry viscera is labor-intensive and prone to errors.
• Automated systems are needed to improve efficiency and accuracy in poultry processing plants.

Purpose of the Study:

• To develop and evaluate a pattern recognition framework for semantic segmentation of poultry organs.
• To replace strenuous manual inspection with an automated, objective method.

Main Methods:

• Utilized a pattern recognition framework for pixel-level multi-class labeling of organs in RGB-D images.
• Extracted features from convolutional neural network (CNN) activation maps.
• Employed a random forest classifier and conditional random fields for refined segmentation.
• Integrated 2D, 3D, and CNN-derived features.

Main Results:

• Achieved a mean Jaccard index of 78.11% across four organ classes.
• Demonstrated improved performance compared to using only basic 2D image features (74.28%).
• The framework successfully segmented organs from 604 RGB-D images of eviscerated poultry viscera.

Conclusions:

• The proposed framework offers a viable automated solution for organ segmentation in poultry processing.
• Combining 2D, 3D, and CNN features significantly enhances segmentation accuracy.
• This approach represents a significant step towards modernizing poultry inspection processes.