초해상도 이미지 생성을 위한 SE Block과 Inception 모듈 기반 ESRGAN 판별자
Discriminator of ESRGAN based on Squeeze-and-Excitation Block and Inception Module for Qualitative Image
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
생성적 적대 신경망(Generative Adversarial Network, GAN)을 활용한 초해상도 기술은 다양한 분야에서 주 목받고 있다. 특히 저해상도 이미지를 고해상도로 변환할 때, 보다 자연스럽고 사실적인 텍스처 복원 능력을 보여준다. 하지만 기존의 초해상도 GAN 모델들은 여전히 고해상도 이미지 생성 시 질감, 가장자리, 패턴 등의 세부적인 특성을 정밀하게 복원하는데 한계를 가지고 있다. 본 연구에서는 초해상도 생성 모델, 특히 Enhanced Super-Resolution Generative Adversarial Network(ESRGAN)의 성능을 개선하기 위해 Inception 모듈과 SE (Squeeze-and-Excitation) Block을 병합한 판별자 구조를 제안한다. 이러한 접근 방법을 통해 모델은 다양한 스케일의 이미지 특징을 보다 효과적으로 학습하고 중요한 정보에 집중할 수 있게 된다. 제안하는 판별자 기반 ESRGAN 모델은 이미지 품질 평가지표에서 PSNR은 29.07dB, SSIM은 0.82, LPIPS는 0.27, NIQE는 5.85를 달성하여 기존 모델 대비 개선되었음을 보인다.
Abstract
GAN models in super-resolution technology have garnered significant attention across various fields. The restoration of more natural and realistic textures is remarkably demonstrated, when converting low-resolution images to high-resol ution counterparts. However, the traditional GAN models for super-resolution still face limitations in precisely reconstr ucting detailed characteristics such as textures, edges, and patterns when generating high-resolution images. This paper proposes a novel discriminator architecture that integrates the Inception module and SE Block to enhance the perform ance of super-resolution generative models, particularly the ESRGAN. This model is able to learn multi-scale image f eatures and focus on critical information more effectively. The proposed model achieves performance improvements ov er the existing ESRGAN model, with a PSNR of 29.07 dB, SSIM of 0.82, LPIPS of 0.27, and NIQE of 5.85 on im age quality evaluation metrics.
참고문헌 (28)
- Z. Wang, "Deep learning for image super-resolution : a survey", IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10), pp.3365~3387 (2020)
- A. Creswell, "Generative Adversarial Network s : An Overview", IEEE Signal Processing Magazine, 35(1), pp.53~65 (2018)
- T.-Y. Lim, "Develop ment of compound eye image quality improvement based on E SRGAN", Journal of the Korea Computer Graphics Society, 30(2), pp.11~19 (2024)
- N. Z. F. N. Azam, "Super re solution with interpolation-based method : a review", IJRAR-Inte rnational Journal of Research and Analytical Reviews, 9(2), pp.168~174 (2022)
- S. Anwar, "A deep journey into super-resolution : A survey", ACM Computing Surveys(CSUR), 53(3), pp.1~34 (2020)
- W. S. Lai, "Dee p Laplacian pyramid networks for fast and accurate super-resol ution", Proceedings of the IEEE Conference on Computer Visi on and Pattern Recognition(CVPR), pp.624~632 (2017)
- Y. Tai, "MemNet : A persiste nt memory network for image restoration", Proceedings of the IEEE International Conference on Computer Vision(ICCV), pp.4539~4547 (2017)
- B. Liu, "A super resolution algorithm based on attention mechanism and SRGAN network", IEEE Access, 9, pp.139138~139145 (2021)
- X. Wang, "ESRGAN : Enhanced super-resolution gener ative adversarial networks", Proceedings of the European Confe rence on Computer Vision (ECCV) Workshops, pp.1~15 (2018)
- Y. Rao, "Deep residual dense network f or single image super-resolution", Electronics, 10(5), pp.555~ (2021)
- N. Rakotonirina, "Further i mproving enhanced super-resolution generative adversarial netw ork", 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2020)
- J. Liang, "Details or Artifacts : A Locally Discriminative Learning Approach to Realistic Image S uper-Resolution", Proceedings of the IEEE, 2022, 5657~5666
- C. Szegedy, "Going deeper with convolutions", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.1~9 (2015)
- X. Ding, "Diverse Branc h Block : Building a Convolution as an Inception-like Unit", Proceedings of the IEEE, 2021, 10886~10895
- A. Song, "Fully Convolutional Networks with Multiscale 3D Filters and Transfer Learning for Change Detecti on in High Spatial Resolution Satellite Images", Remote Sensing, 12(5), pp.799~ (2020)
- J. Hu, "Squeeze-and-excitation netw orks", Proceedings of the IEEE Conference on Computer Visio n and Pattern Recognition (CVPR), pp.7132~7141 (2018)
- X. Cheng, "SESR : Single im age super resolution with recursive squeeze and excitation netw orks", 2018 24th International Conference on Pattern Recogn ition (ICPR), pp.147~152 (2018), IEEE
- J. He, "GCFSR : A Generative and Controllable Face Super Resolution Method Without Facial and GAN Priors", Proceedings of the IEEE, 2022, 1889~1898
- Y. Zhang, "MR Image Super-R esolution With Squeeze and Excitation Reasoning Attention Net work", Proceedings of the IEEE, 2021, 13425~13434
- C. Szegedy, "Ince ption-v4, inception-ResNet and the impact of residual connections on learning", Proceedings of the AAAI Conference on A rtificial Intelligence, 31(1), pp.1~15 (2017)
- A. G. Roy, "Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolution al Networks", Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, pp.421~429 (2018)
- C. Ma, "Uncertainty-Aware GAN for Single Image Super Resolution", Proceedings of the AAAI Conference on Artificial Intelligence, pp.4071~4079 (2024)
- E. Agustsson, "NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp.1122~1131 (2017)
- A. Bulat, "To Learn Image Super-Resolution, Use a GAN to Learn How to Do Image Degradation First", Proceedings of the European Conference on Computer Vision (ECCV), pp.185~200 (2018)