Deep orthogonal transform feature for image denoising
- 주제(키워드) Image denoising , Deep learning
- 발행기관 고려대학교 대학원
- 지도교수 김종옥
- 발행년도 2019
- 학위수여년월 2019. 2
- 유형 Text
- 학위구분 석사
- 학과 대학원 전기전자공학과
- 세부전공 신호 및 멀티미디어전공
- 원문페이지 33 p
- 실제URI http://www.dcollection.net/handler/korea/000000083414
- UCI I804:11009-000000083414
- DOI 10.23186/korea.000000083414.11009.0000927
- 본문언어 영어
- 제출원본 000045978562
초록/요약
In this paper, we solve a classical image denoising problem, which is to remove observed Gaussian noise from noisy images. There have been proposed many methods for image denoising. Recently, Convolutional Neural Networks(CNN) based approaches, such as DnCNN, have shown superior performance. However, it still has a limitation on the denoising performance of the specific image region. To overcome this problem, we propose novel networks for learning Orthogonal Transform Features(OTF) which are used as multi inputs of the deep neural networks and new architectures for orthogonal transform domain. OTF include features from not only noisy images but smoothed images which are noise-reduced once by a denoising method. Our final results show that denoising networks with OTF can generate images which show better PSNR scores and visual quality in the texture regions than the results of state-of-the-art methods.
more목차
Chapter 1. Introduction 1
Chapter 2. Related Work 4
2.1 Orthogonal Domain based Denoising methods 4
2.2 Deep Learning based Denoising methods 5
Chapter 3. Orthogonal Transform Feature 7
3.1 Wavelet Transform Feature 7
3.2 PCA Feature 8
Chapter 4. Proposed Method 10
4.1 Multi Inputs 10
4.2 Channel-wise Feature Extraction 11
4.3 Channel Fusion and Restoration 12
4.4 Training 13
Chapter 5. Experimental Results 15
5.1 Experimental Setting 15
5.2 Visual Comparison 16
5.3 Effect of Smoothed Feature 17
5.4 Effect of Depthwise Separable Convolution 18
Chapter 6. Conclusion 20
Bibliography 21
