An Improved Crucible Spatial Bubble Detection Based on YOLOv5 Fusion Target Tracking

Area of Science:

• Materials Science
• Chemical Engineering
• Computer Vision

Background:

• Existing crucible bubble detection methods are limited to 2D statistics, hindering comprehensive analysis.
• Current techniques struggle with real-time detection and identifying small targets in quartz crucibles.

Purpose of the Study:

• To develop a novel three-dimensional spatial bubble counting method for enhanced quartz crucible analysis.
• To improve the real-time detection capabilities and accuracy of small bubble identification in industrial settings.

Main Methods:

• Acquisition of spatial video images using a digital microscope to create a quartz crucible bubble dataset.
• Modification of the YOLOv5 network by reducing down-sampling depth and incorporating dilated convolution for feature extraction.
• Integration of an effective channel attention network (ECA-Net) and a Kalman filter-Hungarian matching tracking algorithm for precise bubble counting.

Main Results:

• The proposed detector algorithm significantly reduced the missed detection rate of tiny bubbles.
• Average detection precision increased from 96.27% to 98.76%, with a 50% reduction in model weight.
• The system achieved a processing speed of 82 frames per second (FPS), enabling real-time analysis.

Conclusions:

• The developed 3D spatial bubble counting method offers a significant advancement over existing 2D techniques.
• The enhanced YOLOv5 detector and tracking algorithm provide real-time, high-precision bubble detection in quartz crucibles.
• This method is effective for industrial applications requiring accurate monitoring of spatial bubbles.