A Conditional Generative Adversarial Network and Transfer Learning-Oriented Anomaly Classification System for...

Area of Science:

• Materials Science and Engineering
• Artificial Intelligence
• Image Analysis

Background:

• Accurate classification of nanofibers is crucial for quality control in electrospinning.
• Generating sufficient training data for image classification can be challenging and resource-intensive.

Purpose of the Study:

• To develop a generative model for creating synthetic Scanning Electron Microscope (SEM) images of defective and nondefective nanofibers.
• To implement a transfer learning strategy for classifying these SEM images.
• To assess the effectiveness of the proposed system in reducing the need for extensive laboratory experiments.

Main Methods:

• Development of a conditional-Generative Adversarial Network (c-GAN) to generate synthetic SEM images of nanofibers.
• Implementation of a transfer learning approach using a pre-trained Convolutional Neural Network (CNN).
• Training the CNN on synthetic images and validating its performance on real SEM images.

Main Results:

• The transfer-learned CNN achieved an accuracy rate of up to 95.31% in classifying SEM images.
• The generative model successfully produced synthetic images suitable for training classification models.
• The system demonstrated potential for reducing the number of required laboratory electrospinning experiments.

Conclusions:

• The proposed generative model and transfer learning system offer an effective solution for classifying SEM images of nanofibers.
• This approach can significantly decrease the cost and time associated with industrial-scale electrospinning experiments.
• The findings support the adoption of AI-driven methods for quality control in nanofiber production.