A skeleton and neural network-based approach for identifying cosmetic surface flaws.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Automated inspection systems require robust methods for surface flaw identification.
• Existing methods may struggle with variations in workpiece orientation and position.
• Efficient data representation is crucial for real-time defect detection.

Purpose of the Study:

• To introduce a cosmetic surface flaw identification approach invariant to workpiece orientation and position.
• To develop an efficient method in terms of computer memory usage.
• To reduce the computational cost of training and applying neural networks for flaw detection.

Main Methods:

• Characterizing visual binary images by pixel counts in subskeleton iterations.
• Utilizing a modified Zhou skeleton transform with disk-shaped structuring elements.
• Proposing two coding schemes for subskeleton pixel counts, with and without low-pass filtering.
• Employing a supervised neural network trained via backpropagation with simulated patterns.

Main Results:

• The proposed method demonstrates invariance to workpiece orientation and position.
• The coding schemes significantly reduce skeleton image data, saving neural network run time.
• Off-line training with simulated patterns circumvents the need for collecting flawed samples.
• Experimental results with six workpiece shapes validate the approach's effectiveness.

Conclusions:

• The developed approach offers an efficient and robust solution for cosmetic surface flaw identification.
• The method is suitable for automated inspection systems where workpiece orientation varies.
• The use of subskeleton pixel counts and neural networks provides a promising direction for defect detection.