Abstract
In recent years, large pre-trained deep neural networks (DNNs) haverevolutionized the field of computer vision (CV). Although these DNNs have beenshown to be very well suited for general image recognition tasks, applicationin industry is often precluded for three reasons: 1) large pre-trained DNNs arebuilt on hundreds of millions of parameters, making deployment on many devicesimpossible, 2) the underlying dataset for pre-training consists of generalobjects, while industrial cases often consist of very specific objects, such asstructures on solar wafers, 3) potentially biased pre-trained DNNs raise legalissues for companies. As a remedy, we study neural networks for CV that wetrain from scratch. For this purpose, we use a real-world case from a solarwafer manufacturer. We find that our neural networks achieve similarperformances as pre-trained DNNs, even though they consist of far fewerparameters and do not rely on third-party datasets.