Geometric transformation-based data augmentation on defect classification of segmented images of semiconductor materials using a ResNet50 convolutional neural network

The emergence of machine learning (ML) and deep learning (DL) techniques opens a huge opportunity for their implementation in industry. One of the tasks for which these techniques have the greatest potential is visual inspection, since both ML and DL techniques are able to determine relationships between large volumes of data. However, these techniques require large volumes of images that cannot always be captured. As a solution to this, data augmentation techniques are applied. In this work, starting from a highly imbalanced dataset containing images of different semiconductor defects, several datasets are generated using data augmentation techniques based on geometric transformations. A ResNet50 convolutional neural network is then fed with each of the datasets to analyze the effect of data augmentation on its classification performance. The results reveal that, with an adequate number of synthetic images of the minority classes, the F1-score improves by 3.74% over that obtained with the original dataset.