204 0

Model Compression and Acceleration Based on Real-time Style Transfer

Title
Model Compression and Acceleration Based on Real-time Style Transfer
Author
주인걸
Alternative Author(s)
주인걸
Advisor(s)
문영식
Issue Date
2020-08
Publisher
한양대학교
Degree
Master
Abstract
The goal of style transfer algorithms is to render the content of one image using the style of another. Although the models trained by the current image style transfer algorithms have a strong performance, they will occupy massive computing resources due to numerous parameters, high storage and computation costs and other defects. In addition, most of their related applications are PC-based, which makes it difficult to deploy them on restricted hardware platforms (e.g., mobile devices), and causes some strange checkerboard patterns in the generated images. In response to the above problems, this paper based on the real-time style transfer algorithm of Johnson et al. On the premise of ensuring network performance, based on the original algorithm network structure, a more efficient pruned version of densely connected convolution networks used to replace the original deep residual network to achieve the reduction of network parameters, all three original deconvolution layers discarded. The first two layers use the nearest neighbor interpolation and convolution method for upsampling to avoid the occurrence of checkerboard grids. To test the performance of the improved style mode, we used the traditional subjective evaluation method and the image evaluation method (NIMA) based on convolutional neural networks to score the transferred image quality of the style model before and after improvement. Experiments show that our model is much higher than the existing technology visually and model size, the model size is reduced from the original 1065kb to the improved 185kb, the model compression ratio is 5.75. The improved method overcomes the problems of costly storage and calculation of image style transfer model and huge calculation resource consumption. It also effectively avoids the checkerboard artifacts.
URI
https://repository.hanyang.ac.kr/handle/20.500.11754/152780http://hanyang.dcollection.net/common/orgView/200000438201
Appears in Collections:
GRADUATE SCHOOL[S](대학원) > COMPUTER SCIENCE & ENGINEERING(컴퓨터공학과) > Theses (Master)
Files in This Item:
There are no files associated with this item.
Export
RIS (EndNote)
XLS (Excel)
XML


qrcode

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE