DenseNet With Deep Residual Channel-Attention Blocks for Single Image Super Resolution

Dong-Won Jang, Rae-Hong Park; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

Abstract


This paper proposes a DenseNet with deep Residual Channel Attention (DRCA) for single image super resolution. Recent works have shown that skip connections between layers improve the performance of the convolutional neural network such as ResNet and DenseNet. We have interpreted the role of ResNet (feature value refinement by addition) and DenseNet (feature value memory by concatenation). The contribution of the proposed network is dense connections between residual groups rather than convolution layers. In terms of feature value refinement and memory, the proposed method refines the feature values sufficiently (by residual group) and memorizes the refined feature values intermittently (by dense connections between residual groups). Experimental results show that the proposed DRCA (14.2M) achieved better performance than the state-of-the-art methods with fewer parameters.

Related Material


[pdf]
[bibtex]
@InProceedings{Jang_2019_CVPR_Workshops,
author = {Jang, Dong-Won and Park, Rae-Hong},
title = {DenseNet With Deep Residual Channel-Attention Blocks for Single Image Super Resolution},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}