A Standardized PMML Format for Representing Convolutional Neural Networks with Application to Defect Detection

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are widely used in engineering and manufacturing for image processing.
• Challenges exist in managing and distributing trained CNN models due to a lack of standardized representation and poor framework interoperability.

Purpose of the Study:

• To propose a standardized format for CNNs based on Predictive Model Markup Language (PMML).
• To develop a new schema for representing various CNN systems (classification, regression, semantic segmentation).
• To demonstrate practical application and evaluate performance of the proposed standard.

Main Methods:

• A novel standardized schema for CNN representation using PMML was developed.
• A high-performance scoring engine was created to evaluate images and videos against PMML models.
• Benchmarking was conducted on different computational platforms to assess performance.

Main Results:

• The proposed PMML format successfully represented a semantic segmentation model for detecting casting defects in X-ray images.
• The developed scoring engine demonstrated effective evaluation of images and videos.
• Performance benchmarking provided insights into the utility across various platforms.

Conclusions:

• The proposed PMML-based standard addresses the need for standardized CNN representation and improves interoperability.
• The developed schema and scoring engine offer a practical solution for managing and deploying CNN models in industrial applications.
• The study validates the effectiveness and utility of the proposed standard through performance evaluation.