Image Generation Methods Based on Convolutional Neural Networks
| Autor: | LI, PEI-YING, 李培英 |
|---|---|
| Rok vydání: | 2019 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 107 Image generation methods based on convolutional neural networks are studied, including style image generation, super-resolution image generation and fake image generation. In the style image generation, the advantages and disadvantages of style transfer and fast style transfer are analyzed. The VGG-19 network is used to extract the feature maps which have their implicit meaning as a learning basis, it can avoid unnecessary parameter adjustments. The superiority is not only for style transfer, but also for color transfer and even texture transfer tasks. In the super-resolution image generation, a single-image super-resolution using residual convolutional neural network (CNN) is presented. The proposed residual CNN architecture is constructed by using various experiments as auxiliary guideline. The deconvolutional layer is also employed to eliminate the need of bicubic interpolation in the image preprocessing and to speed up the calculations of network. In addition, although only 50000 tiny images are used for training, this application demonstrates effective training and can achieve deep learning effects. Several public benchmarks are tested to show that the proposed network architecture has better performance than the plain convolutional networks without using residual connection. Finally, the fake image generation based on convolutional generative adversarial networks (GAN) is studied. GAN simultaneously trains two CNN model: one is imgae generation, the other is used to evaluate the goodness or badness of image generation. Using the competing relationship between the two networks, the quality of image generation is improved. In the application, two different vehicle data sets are used for training. Analysis of the coding and generation results provides a trick to improve quality. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|