Quality photo sketch with improved deep learning structure
Author(s)
Date Issued
2022
Publisher
IEEE
Related Publication(s)
Proceedings of the 2022 IEEE Region 10 Conference (TENCON)
Start page
838
End page
843
Abstract
Drawing a sketched picture from realistic scene or photo is useful. In this paper, we propose a high-quality sketch generating model using deep convolutional neural network with self-attention structure. In style-transfer investigation, how to balance and retain both information details of input and style are what we want. For sketch drawing, edges or contours are the major components to form a sketch-like image. However, how to choose edges and contours are the major topics for the model to learn. Besides, keeping a small amount of texture and shadow can give a better view of a sketch result. We resolve this problem by proposing an end-to-end jump connection with elementwise multiplication instead of addition to keep texture details of the original input, which gives highlight of edges and contours for a sketch output. Experimental results show that our new design of network surpasses other state-of-the-art models in sketch details.
SFU Affiliated Publication
Yes
Availability at SFU Library
No database links found.