A Multiscale Instance Segmentation Method Based on Cleaning Rubber Ball Images

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Process Monitoring

Background:

• Accurate identification of wear rubber balls in heat exchange equipment is vital for effective descaling.
• Existing methods struggle with real-time segmentation of rubber ball images due to impurities and bubbles.

Purpose of the Study:

• To develop a multi-scale feature fusion real-time instance segmentation model for rubber ball object segmentation.
• To enhance the accuracy and speed of rubber ball identification in industrial cleaning systems.

Main Methods:

• Proposed a model incorporating Pyramid Vision Transformer (PVT) and spatial-reduction attention for improved feature extraction.
• Enhanced feature fusion with an attention mechanism for better representation.
• Modified the prediction head with dynamic convolution and increased upsampling layers for improved mask accuracy.

Main Results:

• Achieved a 4.5% improvement in Dice score, 4.7% in Jaccard coefficient, and 7.73% in mAP on a custom rubber ball dataset.
• Realized a segmentation speed of 33.6 fps with 79.3% segmentation accuracy.
• Demonstrated superior performance compared to DeepMask, Mask R-CNN, BlendMask, SOLOv1, and SOLOv2.

Conclusions:

• The proposed multi-scale feature fusion model effectively addresses challenges in rubber ball image segmentation.
• The integration of PVT and attention mechanisms enhances segmentation accuracy and real-time processing capabilities.
• This model offers a promising solution for improving descaling efficiency in heat exchange equipment.