Lightweight saliency detection method for real-time localization of livestock meat bones

Area of Science:

• Computer Vision
• Robotics
• Image Processing

Background:

• Existing salient object detection (SOD) networks are computationally intensive, limiting their use in robotics.
• Boning robots require efficient and accurate localization of livestock meat bones.

Purpose of the Study:

• To develop a lightweight saliency detection algorithm for real-time livestock bone localization.
• To address the computational limitations of current SOD networks in robotic applications.

Main Methods:

• A lightweight feature extraction network with multi-scale attention was designed.
• Jump connections were integrated into the decoding phase for enhanced feature fusion.
• A residual refinement module was added to optimize target regions and boundaries.

Main Results:

• The proposed method achieves high detection accuracy with 40 times fewer parameters than conventional models.
• Achieved an Fωβ score of 0.699 on a challenging SOD dataset.
• Demonstrated effective livestock bone segmentation on a custom Pig leg X-ray (PLX) dataset.
• Model achieves a detection speed of 5 frames per second on industrial control equipment.

Conclusions:

• The lightweight saliency detection algorithm is suitable for real-time livestock bone localization in boning robots.
• The proposed method offers a significant reduction in parameters without compromising detection accuracy.
• The algorithm shows promise for improving efficiency and applicability in meat processing automation.