Development of an Image Style Conversion Algorithm Using Machine Learning and Parallel Computing Algorithms

The main goal of this work is to increase the dimensionality of images in the stylization algorithm and create opportunities for artists, designers, cinema workers, etc. to effectively stylize and use high-resolution images in their professional field. In this paper, several possible extensions and improvements to the original neural-style transfer algorithm by Gatys et al (2016) were considered. Google Colaboratory, a free cloud platform, was used to generate images, and Google Cloud Platform computing resources were leased for testing stylization process using parallel computing algorithms. As part of the diploma project, an algorithm that can generate high-resolution stylized output images using CUDA, cuDNN, and parallel computing algorithms has been developed. After describing the principles of the algorithm and the testing conducted, the main limitations and shortcomings of the algorithm were identified, and solutions to speed up the calculations and reduce the time of stylization were proposed. Regarding the experimental part of the final year project, the search for the optimal parameters for the developed algorithm was carried out, as well as studies of multi-style transfer algorithms, fast style transfer, photorealistic style transfer, style transfer with preservation of original color, brush size control, and many others. An explanation of the necessary modifications that are sufficient to achieve these effects, and some considerations are suggested for further study. Further research involves improving the developed algorithm, speeding up calculations, optimizing programming code, and adding a variety of new functionality. The algorithms were implemented in the Python 3.7 programming language in conjunction with the use of the machine learning framework PyTorch 1.4, created by Facebook, to perform calculations on a neural network. This work contains 115 pages, 77 figures, 7 tables, 11 formulas, 7 applications and 94 sources. Keywords: computer system for image stylization, machine learning, deep learning, neural networks, convolutional neural networks, high-resolution image generation, neural style transfer algorithm, parallel computing, cloud computing, PyTorch