Highly constrained neural networks for industrial quality control.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Artificial neural networks (ANNs) typically require large datasets for training.
• Applying ANNs to industrial quality control is challenging due to limited available examples.
• Domain-specific knowledge, like spatial invariances, is often not effectively utilized in standard ANN architectures.

Purpose of the Study:

• To develop techniques for embedding domain-specific spatial invariances into highly-constrained neural networks.
• To reduce the number of weights requiring training, enabling ANN application with limited data.
• To apply and evaluate this methodology for the optical inspection of machined parts.

Main Methods:

• Investigated techniques for embedding spatial invariances into neural network architectures.
• Developed a constrained neural network specifically for optical inspection tasks.
• Trained and evaluated the network using images of machined parts to detect defects.

Main Results:

• The proposed methodology significantly reduces the number of weights needed for training.
• The developed neural network successfully identifies defective machined parts.
• Achieved comparable accuracy to standard classifiers with substantially less training data and faster development time.

Conclusions:

• Embedding domain-specific spatial invariances is a viable strategy for improving neural network efficiency.
• This approach offers a promising solution for automated quality control in industrial settings.
• The methodology provides a balance of accuracy and reduced software development time for defect detection.