A Unified Framework for Super-Resolution Based on Segmentation-Prior and Self-Attention

Convolutional Neural Network (CNN) is intensively applied to super-resolution (SR) task because of its superior performance. However, the problem of SR task is still challenging due to the lack of prior knowledge and small receptive field of CNN. We propose a unified framework for single image SR ba...

Full description

Saved in:
Bibliographic Details
Published in:2022 IEEE 2nd International Conference on Data Science and Computer Application (ICDSCA) pp. 80 - 84
Main Authors: Hu, Xinyi, Zhang, Yuxin, Yang, Dongxiao
Format: Conference Proceeding
Language:English
Published: IEEE 28.10.2022
Subjects:
Computer applications
Convolutional neural networks
Data science
GAN
Generative adversarial networks
Image segmentation
Self-Attention
Semantic Segmentation
Super-Resolution
Superresolution
Task analysis
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Convolutional Neural Network (CNN) is intensively applied to super-resolution (SR) task because of its superior performance. However, the problem of SR task is still challenging due to the lack of prior knowledge and small receptive field of CNN. We propose a unified framework for single image SR based on segmentation-prior and self-attention, named Segmentation-Prior Self-Attention Generative Adversarial Network (SPSAGAN). This combination is led by a carefully designed weighted addition to balance the influence of feature and segmentation attentions. Thus, the SPSAGAN can emphasize textures in the same segmentation category and meanwhile focus on the long-distance feature relationship. Extensive experiments show that SPSAGAN can generate more realistic and visually pleasing textures compared to state-of-the-art SFTGAN [1] and ESRGAN [2] on OST and BSD100 datasets
DOI:10.1109/ICDSCA56264.2022.9988090