Publication:
Enhanced dense space attention network for single image super-resolution

Loading...
Thumbnail Image
Date
2022-02-01
Authors
Ooi, Yoong Khang
Journal Title
Journal ISSN
Volume Title
Publisher
Research Projects
Organizational Units
Journal Issue
Abstract
The development of deep learning has received much attention in the single image super-resolution reconstruction application. The first convolutional neural network (CNN)-based image super-resolution model was the Super-resolution Convolutional Neural Network (SRCNN). Since then, many researchers have put efforts into developing the CNN-based model for image super-resolution to improve the accuracy and reduce the running time of the model. Until today, some models still suffer from the vanishing-gradient problem and rely on a large number of layers that result in a long-running time. Therefore, an enhanced dense space attention network (EDSAN) model is proposed to overcome the problems. The objectives of this projectare to improve the accuracy of the model and reduce the running time by reducing the number of layers required. This project developed a Local Wider Dense Space Attention Block (LWDSAB) in the EDSAN model that adopted a dense connection and residual network to utilize all the features to correlate the low-level feature and high-level feature. Besides, the convolutional block attention module (CBAM) layer and multiscale block (MSB) are deployed in the model to reduce the running timewithout affecting the model’s performance. The model is evaluated through peak signal-to-noise ratio (PNSR) and structural similarity index measure (SSIM) metrics. For state-of-the-art comparison purposes, a total of 4 recent models were taken for results benchmarking. Besides, a total of 4 different types of datasets will be used forperformance evaluation. Results show EDSAN made a different amount of improvement respective to different datasets and different scale factors when compared to different models. EDSAN model performed the best for the Set5 dataset. The greatest improvement made for the PSNR value was 8.96% relative to the DRDN model at a scale factor of 4. For other datasets, EDSAN showed a positive result in the PSNR value at a scale factor of 2 and 3. Although EDSAN is not the top performer at a scale factor of 4, the percentage of the PSNR improvement is more significant than those models that outperform the EDSAN. In terms of SSIM, the EDSAN model showed a positive result for all datasets and all scale factors compared to other models. The highest achievement made was 10.29% relative to the DRDN model for the Urban100 dataset at a scale factor of 4. In conclusion, EDSAN successfully solved the vanishing-gradient problem and long-running time issue.
Description
Keywords
Citation