Improving Recognition of Defective Epoxy Images in Integrated Circuit Manufacturing by Data Augmentation

Area of Science:

• Artificial Intelligence
• Materials Science
• Manufacturing Engineering

Background:

• Vision-based inspection is crucial for die attachment in integrated circuit (IC) manufacturing.
• Accurate recognition of defective epoxy drops is essential for quality control.
• Limited availability of defective epoxy drop images poses a challenge for model training.

Purpose of the Study:

• To evaluate deep neural network models for recognizing defective epoxy drop images.
• To investigate the impact of data augmentation on recognition accuracy.
• To compare supervised and unsupervised learning approaches for this task.

Main Methods:

• Utilized supervised models: Autoencoder (AE) + Multi-layer Perceptron (MLP), and VGG16.
• Employed unsupervised models: AE + k-means clustering, and VGG16 + k-means clustering.
• Generated synthetic defective epoxy drop images using a CycleGAN generative network for data augmentation.

Main Results:

• Data augmentation substantially improved recognition accuracy across all models.
• VGG16-based models (supervised and unsupervised) achieved perfect or near-perfect accuracy with data augmentation.
• Supervised AE+MLP accuracy improved by 47%, unsupervised AE+Kmeans by 37% due to augmentation.

Conclusions:

• Data augmentation is highly effective for improving deep learning-based epoxy drop defect recognition.
• VGG16 models demonstrate superior performance for vision-based die attachment inspection.
• The study highlights the potential of generative networks for enhancing training datasets in manufacturing quality control.