An industrial carbon block instance segmentation algorithm based on improved YOLOv8

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Machine vision is key for industrial intelligent applications, particularly in automated carbon block cleaning.
• Accurate carbon block type recognition and center point localization are critical but challenging for existing instance segmentation algorithms.

Purpose of the Study:

• To develop an improved instance segmentation algorithm (YOLOv8-HDSA) for accurate industrial carbon block recognition and edge segmentation.
• To enhance feature representation and fusion capabilities for better performance in industrial settings.

Main Methods:

• Proposed YOLOv8-HDSA algorithm featuring a Selective Reinforcement Feature Fusion Module (SRFF) using Hadamard product and dilated convolution.
• Incorporated a convolutional self-attention mechanism with residual structure in the head for improved feature extraction.
• Introduced Focaler-IoU as the loss function to optimize regression performance.

Main Results:

• YOLOv8-HDSA demonstrated significant improvements on real industrial datasets.
• Achieved a 7.2% increase in average carbon block recognition accuracy.
• Improved carbon block edge segmentation accuracy by 3.8%.

Conclusions:

• The proposed YOLOv8-HDSA algorithm effectively addresses limitations of existing methods for industrial carbon block instance segmentation.
• The enhancements in feature fusion, attention mechanisms, and loss function contribute to superior recognition and segmentation accuracy.