An Improved Mask R-CNN Micro-Crack Detection Model for the Surface of Metal Structural Parts

Area of Science:

• Engineering
• Materials Science
• Computer Vision

Background:

• Micro-crack detection is crucial for critical equipment health monitoring and ensuring operational stability.
• Conventional target detection models exhibit low accuracy in identifying micro-cracks on metal structures.

Purpose of the Study:

• To enhance the accuracy and performance of micro-crack detection using an optimized Mask R-CNN model.
• To address limitations in current models for detecting micro-cracks on metal structural parts.

Main Methods:

• Developed a dedicated micro-cracks dataset.
• Improved the Feature Pyramid Network (FPN) with a bottom-up feature fusion path.
• Integrated deformable convolution kernels and attention mechanisms into ResNet for efficient feature extraction.
• Modified the loss function to optimize network training and convergence.

Main Results:

• All proposed improvement schemes individually enhanced the original Mask R-CNN performance.
• The integrated approach yielded significant improvements in micro-crack recognition, classification, and positioning.
• Ablation experiments validated the effectiveness of each modification.

Conclusions:

• The optimized Mask R-CNN demonstrates superior performance for micro-crack detection.
• The proposed enhancements are rational and feasible for improving equipment health monitoring.
• This work provides a robust method for accurate micro-crack identification in structural components.