DOI QR코드

DOI QR Code

A Performance Comparison of Super Resolution Model with Different Activation Functions

활성함수 변화에 따른 초해상화 모델 성능 비교

  • Received : 2020.07.07
  • Accepted : 2020.07.28
  • Published : 2020.10.31

Abstract

The ReLU(Rectified Linear Unit) function has been dominantly used as a standard activation function in most deep artificial neural network models since it was proposed. Later, Leaky ReLU, Swish, and Mish activation functions were presented to replace ReLU, which showed improved performance over existing ReLU function in image classification task. Therefore, we recognized the need to experiment with whether performance improvements could be achieved by replacing the RELU with other activation functions in the super resolution task. In this paper, the performance was compared by changing the activation functions in EDSR model, which showed stable performance in the super resolution task. As a result, in experiments conducted with changing the activation function of EDSR, when the resolution was converted to double, the existing activation function, ReLU, showed similar or higher performance than the other activation functions used in the experiment. When the resolution was converted to four times, Leaky ReLU and Swish function showed slightly improved performance over ReLU. PSNR and SSIM, which can quantitatively evaluate the quality of images, were able to identify average performance improvements of 0.06%, 0.05% when using Leaky ReLU, and average performance improvements of 0.06% and 0.03% when using Swish. When the resolution is converted to eight times, the Mish function shows a slight average performance improvement over the ReLU. Using Mish, PSNR and SSIM were able to identify an average of 0.06% and 0.02% performance improvement over the RELU. In conclusion, Leaky ReLU and Swish showed improved performance compared to ReLU for super resolution that converts resolution four times and Mish showed improved performance compared to ReLU for super resolution that converts resolution eight times. In future study, we should conduct comparative experiments to replace activation functions with Leaky ReLU, Swish and Mish to improve performance in other super resolution models.

ReLU(Rectified Linear Unit) 함수는 제안된 이후로 대부분의 깊은 인공신경망 모델들에서 표준 활성함수로써 지배적으로 사용되었다. 이후에 ReLU를 대체하기 위해 Leaky ReLU, Swish, Mish 활성함수가 제시되었는데, 이들은 영상 분류 과업에서 기존 ReLU 함수 보다 향상된 성능을 보였다. 따라서 초해상화(Super Resolution) 과업에서도 ReLU를 다른 활성함수들로 대체하여 성능 향상을 얻을 수 있는지 실험해볼 필요성을 인지하였다. 본 연구에서는 초해상화 과업에서 안정적인 성능을 보이는 EDSR(Enhanced Deep Super-Resolution Network) 모델에 활성함수들을 변경하면서 성능을 비교하였다. 결과적으로 EDSR의 활성함수를 변경하면서 진행한 실험에서 해상도를 2배로 변환하는 경우, 기존 활성함수인 ReLU가 실험에 사용된 다른 활성함수들 보다 비슷하거나 높은 성능을 보였다. 해상도를 4배로 변환하는 경우에서는 Leaky ReLU와 Swish 함수가 기존 ReLU 함수 대비 다소 향상된 성능을 보임을 확인하였다. Leaky ReLU를 사용했을 때 기존 ReLU보다 영상의 품질을 정량적으로 평가할 수 있는 PSNR과 SSIM 평가지표가 평균 0.06%, 0.05%, Swish를 사용했을 때는 평균 0.06%, 0.03%의 성능 향상을 확인할 수 있었다. 또한 해상도를 8배로 변환하는 경우에서는 Mish 함수가 기존 ReLU 함수 대비 다소 향상된 성능을 보임을 확인하였다. Mish를 사용했을 때 기존 ReLU보다 PSNR과 SSIM 평가지표가 평균 0.06%, 0.02%의 성능 향상을 확인할 수 있었다. 결론적으로 해상도를 4배로 변환하는 초해상화의 경우는 Leaky ReLU와 Swish가, 해상도를 8배로 변환하는 초해상화의 경우는 Mish가 ReLU 대비 향상된 성능을 보였다. 향후 연구에서는 다른 초해상화 모델에서도 성능 향상을 위해 활성함수를 Leaky ReLU, Swish, Mish로 대체하는 비교실험을 수행하는 것도 필요하다고 판단된다.

Keywords

References

  1. Saeed Anwar, Salman Khan, and Nick Barnes, "A Deep Journey into Super-resolution: A Survey," arXiv:1904.07523, 1, 2019.
  2. B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee, "Enhanced Deep Residual Networks for Single Image Super-Resolution," In Conference on Computer Vision and Pattern Recognition, 2017, 1.
  3. R. Timofte, E. Agustsson, L. Van Gool, M.-H. Yang, L. Zhang, B. Lim, S. Son, H. W. Kim, et al., "NTIRE 2017 Challenge on Single Image Super-Resolution: Methods and Results," In Conference on Computer Vision and Pattern Recognition, 2017, 1,2,4,6,7,8.
  4. C. Ledig, L. Thesis, F. Huszar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, et al, "Photo-realistic Single Image Super-resolution using a Generative Adversarial network," arXiv:1609.040802,1,2,3,4,5,6,7, 2017.
  5. K. He, X. Zhang, S. Ren, and J. Sun, "Deep Residual Learning for Image Recognition," In Conference on Computer Vision and Pattern Recognition, 2016, 3.
  6. Nair, Vinod and Hinton, Geoffrey E, "Rectified Linear units Improve Restricted Boltzmann Machines," In International Conference on Machine Learning, 2010, pp.807-814.
  7. Bing Xu, Naiyan Wang, Tianqi Chen, Mu Li, "Empirical Evaluation of Rectified Activations in Convolutional Network," arXiv:1505.00853, 1, 2015.
  8. Prajit Ramachandran, Barret Zoph, and Quoc V. Le, "Swish: a Self-gated Activation Function," arXiv:1710.05941 7, 1, 2017.
  9. Diganta Misra, "Mish: A Self Regularized Non-Monotonic Neu ral Activation Function," arXiv:1908.08681, 1, 2019.
  10. M. Bevilacqua, A. Roumy, C. Guillemot, and M. L. Alberti-Morel, "Low-complexity Single-image Super-resolution Based on Nonnegative Neighbor Embedding," In British Machine Vision Conference, 2012, 2,4.
  11. R. Zeyde, M. Elad, and M. Protter, "On Single Image Scale-up using Sparse-representations," In Proceedings of the International Conference on Curves and Surfaces, 2010, 2,4.
  12. D. Martin, C. Fowlkes, D. Tal, and J. Malik, "A Database of Human Segmented Natural Images and its Application to Evaluating Segmentation Algorithms and Measuring Ecological Statistics," In International Conference on Computer Vision, 2001, 4.
  13. J.-B. Huang, A. Singh, and N. Ahuja, "Single Image Super-resolution from Transformed Self-exemplars," In Conference on Computer Vision and Pattern Recognition, 2015, 2,4,6.