Comparison between standard and GAN-based data augmentation to improve chip pick classification

Abstract: Pick-and-place machines are used to assemble printed circuit boards, but they sometimes fail to pick up components correctly. Images of the chips once picked up are classified by a neural network to determine whether the process was successful. However the number of images in the dataset with certain faulty orientations is suspected to be insufficient to train the classifier satisfyingly, but gathering more data would be both troublesome and costly. Data augmentation techniques make it possible to train a classifier with extra samples, which may either be mere altered real images or completely new images created by a neural network. For the latter, the most common approach is to turn to Generative Adversarial Networks (GANs), which use two competing neural networks to try to produce the most realistic images possible. This work explores how to improve classification performance by tuning and comparing standard and GAN data augmentation, with GANs either generating images out of nothing or out of pre-existing simulated images. Although the two techniques based on GANs did not outperform standard data augmentation, significant improvements were obtained with these three methods.