A research team led by Prof. YANG Zhijia from the Shenyang Institutes of Automation of the Chinese Academy of Sciences has developed image augmentation method based on adversarial generative network (GAN) for the limited number of samples, which was used for detecting bubbles in the photoresist during semiconductor manufacture.

The results were published in Engineering Applications of Artificial Intelligence (EAAI).

In the actual spin-coating procedure of semiconductor manufacture, collecting high-quality photoresist images for extended periods of time is very expensive and labor intensive, and the bubble defects are not present all the time.

In this study, the researchers designed a GAN to generate images containing bubble defects to augment original dataset. What’s more, the researchers utilized the augment data to train a well-performed neural network for detecting bubbles.

They adopted an autoencoder as the improved mapping network, which took the original image and the position code as input and the same image after vertical shifting as output. The different position codes corresponded to different vertical shifting lengths, finally generating various bubble images.

The results show that the proposed GAN can generate samples that are more variable and are not distinguishable directly from real images.

This work is supported by the Special Fund for Local Science and Technology Development of Central Guidance, the National Key Research and Development Program of China, and Nature Science Foundation of Liaoning province.

The network structure of improved GAN.

The generated image v.s. the original images.