A deep learning-based demosaicking and denoising method with wavelet channel attention
웨이블릿 채널 어텐션을 이용한 딥러닝 기반 디모자이킹 및 노이즈 제거 방법
- 주제(키워드) Image restoration , image reconstruction , Demosaicking , Denoising , Deep learning
- 주제(DDC) 621.381
- 발행기관 아주대학교
- 지도교수 선우명훈
- 발행년도 2022
- 학위수여년월 2022. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/ajou/000000031405
- 본문언어 한국어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Demosaicking and denoising are very important tasks among various processing steps of Image signal processor (ISP) because they are essential and process raw data obtained from sensors. Convolutional Neural Network (CNN) based Joint Demosaicking and denoising methods simultaneously process Demosaicking, and denoising using a single network are also proposed. Although existing CNN based methods showed excellent performance, the high-frequency details of the image were still not well restored, and a network structure for learning focused on that part was not be proposed. To solve this problem, we proposed a network structure that learns features from multi-resolution feature maps after downsampling without data loss by applying Discrete Wavelet Transform (DWT) to CNN. Moreover, this paper proposed a network structure that pays attention to the high frequency of an image using the channel attention technique and a loss function that reduces the loss in the frequency domain. The proposed methods are high-efficiency methods that show high-performance improvement compared to required parameters or memory. In addition, the performance of the proposed method achieved the highest PSRN and SSIM compared with existing methods and showed a result of well reconstructing high-frequency details such as edges when comparing reconstructed images.
more목차
I Introduction 1
II Related Works 3
A Demosaicking and Denoising 3
B DWT for Image Reconstruction 4
C Channel Attention 4
III Proposed Method 6
A Proposed Network Structure 6
B Wavelet Attention Block 8
C Loss Function 9
IV Experimental Results 11
A Details of Network Training 11
B Comparison of Overall Results 12
C Ablation Study of Proposed Network 15
V Conclusion 17
Bibliography 18