The thesis presents an application designed to virtually fit different hairstyles to an image using computer vision tools. It uses two methods of segmentation, which divide the input image into three separate semantic regions (hair, face and background) and two methods of inpainting, which fill the holes left in the process of combining hair with faces.
The neural networks FCN-8s and U-Net used in this work, are trained and validated on the LFW dataset. The results indicate, that U-Net outperforms FCN-8s in hair segmentation tasks, where it achieved a higher average per-pixel accuracy of 0.910, where FCN-8s’ average per-pixel accuracy was 0.902, respectively. The F-score results, where U-Net achieved better accuracy across all classes, further confirms the previous observations. We used a deep generative neural network and the Thelea method, which fills holes according to its neighboring pixels, for inpainting. The Thelea method achieved a slightly better average score of 4.214 in the human rating test, as opposed to 3.429 of the deep generative neural network.
Chapter 1 provides an introduction to the thesis. The first part of Chapter 2 presents a short overview of the fundamental elements in deep learning, whereas the second part explores related work. Chapter 3 explains the methodology behind the application. Chapter 4 presents the results, and validates the segmentation and inpainting methods. Chapter 5 presents the conclusion and further work.
|
