Institutional Repository of Vilnius University
Pašalinių objektų šalinimo nuotraukose giliaisiais neuroniniais tinklais tyrimas /
Abstract
dc:descriptionThe aim of this work is to explore ways of removing and inpainting objects, improving models and developing a mobile application for easy use. The inpainting techniques investigated are divided into two groups: traditional algorithms and deep neural networks. Among the traditional algorithms, the following are described: the use of a diffusion kernel, automatic object segmentation with statistical region-growing, texture synthesis, and the use of Navier-Stokes partial differential equations. Neural networks described are: "Wasserstein GAN-GP" generative adversarial network with contextual attention, state-of-the-art "Segment Anything" object segmentation model, state-of-the-art LaMa convolutional inpainting model, and the combination of the latter two neural networks into the "Inpaint Anything" system. Ways to improve neural network models are explored: network pruning, quantization, use of Huffman coding. A mobile application is also developed, which allows to select one object to be inpainted by marking it with a dot and then choosing the most appropriate masking. The application uses two traditional algorithms and three neural networks, one of which is an object segmentation model. During the development of this application, resource restrictions were encountered, so it was decided to use smaller and less well trained neural networks and to reduce the dimensions of the input image roughly three times. Both modifications allowed "Wasserstein GAN-GP" to be used for inpainting. It was also decided to run the LaMa model with the central processing unit (CPU), as it did not have enough resources with the full-size photos running with the video card. This had a significant impact on the speed of the mobile application, as it took more than five minutes to inpaint with the neural networks.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Styra, Patrikas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD192846163&prefLang=en_US