Page 196 - Kaleidoscope Academic Conference Proceedings 2020

2020 ITU Kaleidoscope Academic Conference

Table 4 – Effects of each block in SR [7] Y. L. Zhang et al., “Image Super-Resolution Using Very Deep
Residual Channel Attention Networks,” Proceedings of the European
Baseline w/o RFS w/o SC L2 L2 + L1 Conference on Computer Vision (ECCV). 2018: 286-301. 1, 3, 5, 7
PNSR 32.20 32.15 32.10 32.17 32.19
[8] T. Dai, J. Cai, Y. Zhang, S. Xia and L. Zhang, "Second-Order
Without RFS and only using shuffle and data augment via Attention Network for Single Image Super-Resolution," 2019
IEEE/CVF Conference on Computer Vision and Pattern Recognition
rotation and flip for sampling, there’s a 0.05dB quality (CVPR), Long Beach, CA, USA, 2019, pp. 11057-11066, doi:
degradation (32.15dB vs. 32.20dB), which proves that RFS 10.1109/CVPR.2019.01132.
can significantly improve the HR quality.
[9] Q. Huang, D. Yang, P. Wu, H. Qu, J. Yi and D. Metaxas, "MRI
Reconstruction Via Cascaded Channel-Wise Attention
If we remove SC between the inner FMG output and scaling Network," 2019 IEEE 16th International Symposium on Biomedical
part, the degradation is worse which shows the importance Imaging (ISBI 2019), Venice, Italy, 2019, pp. 1622-1626, doi:
of SC. 10.1109/ISBI.2019.8759423.
[10] K. He, X. Zhang, S. Ren and J. Sun, “Deep Residual Learning for
The last two columns show the effectiveness of hybrid Image Recognition,” 2016 IEEE Conference on Computer Vision and
L1+L2+Lp norm. As can be seen, L1 + L2 is better than L1 Pattern Recognition (CVPR), Las Vegas, NV, 2016, pp. 770-778, doi:
alone but less superior than all three norms combined. 10.1109/CVPR.2016.90.
[11] Zhou F, Li X, Li Z. High-frequency details enhancing DenseNet for
Table 4 effectively illustrates the three imbalances super-resolution [J]. Neurocomputing, 2018, 290: 34-42. 1, 5
mentioned in this paper, and the improvement of PSNR also
shows that the proposed method can alleviate these defects. [12] Y. Zhang, Y. Tian, Y. Kong, B. Zhong and Y. Fu, “Residual Dense
Network for Image Super-Resolution,” 2018 IEEE/CVF Conference
on Computer Vision and Pattern Recognition, Salt Lake City, UT,
5. CONCLUSION 2018, pp. 2472-2481, doi: 10.1109/CVPR.2018.00262.
In this paper, we show the imbalance issues in SISR which [13] X. T. Wang et al., “ESRGAN: Enhanced Super-Resolution Generative
includes sample imbalance, feature imbalance, and object Adversarial Networks”, The European Conference on Computer
Vision (ECCV) Workshops, 2018.
function imbalance. To tackle these imbalance problems, a
balanced SR framework is proposed which features a novel [14] J. Cai, H. Zeng, H. Yong, Z. Cao and L. Zhang, "Toward Real-World
random sampling algorithm during training, feature Single Image Super-Resolution: A New Benchmark and a New
extraction group structure, as well as an Lp object function. Model," 2019 IEEE/CVF International Conference on Computer
Vision (ICCV), Seoul, Korea (South), 2019, pp. 3086-3095, doi:
The proposed BSR significantly improves the SR 10.1109/ICCV.2019.00318.
performance. However, not all the imbalance issues in SR
have been studied thoroughly and there’s still lots of space [15] J. Pang, K. Chen, J. Shi, H. Feng, W. Ouyang and D. Lin, "Libra R-
for further improvements, such as how to improve efficiency CNN: Towards Balanced Learning for Object Detection," 2019
IEEE/CVF Conference on Computer Vision and Pattern Recognition
of multi-scale features, which would be our future research. (CVPR), Long Beach, CA, USA, 2019, pp. 821-830, doi:
10.1109/CVPR.2019.00091.
REFERENCES
[16] Y. Cui, M. Jia, T. Lin, Y. Song and S. Belongie, "Class-Balanced Loss
[1] K. Nasrollashi, T. B. Moeslund, “Super-resolution: A Comprehensive Based on Effective Number of Samples," 2019 IEEE/CVF Conference
Survey”, Machine Vision and Applications, 2014, pp. 1423 - 1468. on Computer Vision and Pattern Recognition (CVPR), Long Beach,
CA, USA, 2019, pp. 9260-9269, doi: 10.1109/CVPR.2019.00949.
[2] L. Wang and G. Jeon, "Bayer Pattern CFA Demosaicking Based on
Multi-Directional Weighted Interpolation and Guided Filter," in IEEE [17] K. Oksuz, B. C. Cam, S. Kalkan and E. Akbas, "Imbalance Problems
Signal Processing Letters, vol. 22, no. 11, pp. 2083-2087, Nov. 2015, in Object Detection: A Review," in IEEE Transactions on Pattern
doi: 10.1109/LSP.2015.2458934. Analysis and Machine Intelligence, do
10.1109/TPAMI.2020.2981890.
[3] J. Tarquino, A. Rueda and E. Romero, "Shearlet-based sparse
representation for super-resolution in diffusion weighted imaging [18] P. Adam, G. Sam, C. Soumith, C. Gregory, Y. Edward, D. Zachary, L.
(DWI)," 2014 IEEE International Conference on Image Processing Zeming, D. Alban, A. Luca , and L. Adam. Automatic differentiation
(ICIP), Paris, 2014, pp. 3897-3900, doi: 10.1109/ICIP.2014.7025791. in pytorch. 2017.

[4] W. Yang et. al., “Deep learning for single image super-resolution: a [19] E. Agustsson and R. Timofte, "NTIRE 2017 Challenge on Single
brief review”, IEEE Transactions on Multimedia, Issue 12, pp. 3106 – Image Super-Resolution: Dataset and Study," 2017 IEEE Conference
3121, Dec. 2019. on Computer Vision and Pattern Recognition Workshops (CVPRW),
Honolulu, HI, 2017, pp. 1122-1131, doi: 10.1109/CVPRW.2017.150.
[5] C. Dong, C. C. Loy, K. He and X. Tang, “Image Super-Resolution
Using Deep Convolutional Networks,” in IEEE Transactions on [20] J. K. Pant, W. Lu and A. Antoniou, "New Improved Algorithms for
Pattern Analysis and Machine Intelligence, vol. 38, no. 2, pp. 295-307, Compressive Sensing Based on lp Norm," in IEEE Transactions on
1 Feb. 2016, doi: 10.1109/TPAMI.2015.2439281. Circuits and Systems II: Express Briefs, vol. 61, no. 3, pp. 198-202,
March 2014, doi: 10.1109/TCSII.2013.2296133.
[6] J. Kim, J. K. Lee and K. M. Lee, “Accurate Image Super-Resolution
Using Very Deep Convolutional Networks,” 2016 IEEE Conference [21] M. Kloft, U. Brefeld, S. Sonnenburg, Lp-norm multiple kernel
on Computer Vision and Pattern Recognition (CVPR), Las Vegas, learning[J]. Journal of Machine Learning Research, 2011, 12(Mar):
NV, 2016, pp. 1646-1654, doi: 10.1109/CVPR.2016.182. 953-997. 4

– 138 –