Image Augmentation with Neural Style Transfer
The amount of training data is of crucial importance for the performance of machine learning, and especially deep learning models. It is one of the most important factors that determine whether the developed model is effective or not. When the quantity of
- PDF / 766,422 Bytes
- 13 Pages / 439.37 x 666.142 pts Page_size
- 15 Downloads / 219 Views
2
Netcetera, Skopje, North Macedonia [email protected] FCSE, Ss. Cyril and Methodius University, Skopje, North Macedonia
Abstract. The amount of training data is of crucial importance for the performance of machine learning, and especially deep learning models. It is one of the most important factors that determine whether the developed model is effective or not. When the quantity of training data for a computer vision problem is insufficient, various data augmentation techniques are used to artificially extend the training dataset with samples that retain the natural distribution of the original data. This paper proposes and evaluates a deep learning model that will be used for image augmentation. A complex deep neural network makes use of transfer learning in order to learn the characteristics of the content and style of the training images, create random style embeddings via learned multivariate normal distribution, and ultimately generate images to extend the original dataset. The model is trained on two datasets which are frequently used in computer vision: ImageNet and Painter by Numbers (PBN). Afterwards, the model is used to generate new images from the CIFAR-100 and Tiny-ImageNet-200 datasets. The performance of the augmentation model is evaluated by a separate convolutional neural network. The evaluation model is trained on the combined dataset, consisting of both, the original and augmented images, and then compared to the performance of the same model trained on the original datasets. Keywords: Image augmentation Neural style transfer Computer Vision Convolutional Neural Networks (CNNs) Deep learning
1 Introduction In the current era of deep learning, there is one thing that can always improve a developed model, and that is more data. On one hand, deep learning models are becoming more accurate than every other carefully developed and hand-designed machine learning method, but on the other hand, they also need much more data. Having a small dataset is one of the biggest setbacks in computer vision projects, especially those with a deep learning approach. This paper focuses on exploring image manipulation through neural style transfer, and adopts an approach for randomizing style, in order to achieve arbitrary image augmentation. Neural style transfer is a technique for reconstructing images by changing their style. It all started when Gatys et al. (2015) showed the possibility of using convolutional neural networks for transforming images, such that they are altered by applying © Springer Nature Switzerland AG 2019 S. Gievska and G. Madjarov (Eds.): ICT Innovations 2019, CCIS 1110, pp. 212–224, 2019. https://doi.org/10.1007/978-3-030-33110-8_18
Image Augmentation with Neural Style Transfer
213
styles of other chosen images, whilst preserving their content [2]. A steady progress has been made in a number of research studies [1, 3, 5, 15] since the original idea was proposed. The current state-of-the-art models can generate a new image based on content and style input images in
Data Loading...
