東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁 到查詢結果 [ null ]

切換: 標籤 | MARC模式 | ISBD

Self-supervised Deep Learning for Im...

Zheng, Huan,

FindBook

Google Book

Amazon

博客來

Self-supervised Deep Learning for Image Denoising and Beyond /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Self-supervised Deep Learning for Image Denoising and Beyond // Huan Zheng.
作者:	Zheng, Huan,
面頁冊數:	1 electronic resource (116 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-06, Section: B.
Contained By:	Dissertations Abstracts International86-06B.
標題:	Tomography. -
電子資源:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31777867
ISBN:	9798346780014

Self-supervised Deep Learning for Image Denoising and Beyond /
Zheng, Huan,

Self-supervised Deep Learning for Image Denoising and Beyond /Huan Zheng. - 1 electronic resource (116 pages)

Source: Dissertations Abstracts International, Volume: 86-06, Section: B.

Images are central to the field of computer vision. However, their quality can be significantly compromised by the inevitable degradations introduced during capturing and transmission, leading to inaccuracies in subsequent analyses. Noise is one key issue existing in most degradations. This has motivated the development of robust and efficient denoising techniques. Of late, supervised deep learning methods, which learn a mapping from a noisy image to its clean version using extensive external datasets, have become prominent. However, these methods may perform poorly due to disparities between the training and testing datasets. Moreover, collecting a sizeable training datasets can be challenging under certain circumstances.Recently, there is increasing interest in unsupervised learning-based methods which relax the need for a paired noisy-clean dataset. In this thesis, we initially revisit the innovative Noise2Noise work, which trains a neural network (NN) with only paired independent noisy images. However, collecting such paired datasets still remain a daunting task. Motivated by this, we propose a novel unsupervised approach, Recorrupted to Recorrupted (R2R), which requires only unpaired noisy images for image denoising tasks. Assuming Gaussian noise, we proved that our R2R training is statistically equivalent to its supervised counterpart. Nevertheless, such Gaussian noise assumption often fails in real-world scenarios and we further extended our R2R for general noise distributions, culminating in the Extended R2R (ER2R) method. We validated ER2R's efficacy on video denoising tasks, using a specific NN architecture to efficiently and effectively exploit temporal redundancy. Notably, ER2R can be trained using only a single test video, thus negating the need for external training datasets. This approach meets the same criteria as traditional methods, but with significantly improved performance. Furthermore, we extended our ER2R method to inverse problems, specifically focusing on the non-blind deconvolution problem, and proposed a kernel-perturbation-based inference scheme for managing kernel errors in real-world scenarios. In conclusion, we propose an unsupervised denoising framework that is applicable to many problems including but not limited to image(video) denoising and linear image inverse problems. Compared to supervised learning and other unsupervised learning methods, our methods either offer superior performance or eliminate the need for paired training datasets while still yielding comparable results.

English

ISBN: 9798346780014Subjects--Topical Terms:

836553
Tomography.

Self-supervised Deep Learning for Image Denoising and Beyond /
LDR:03754nmm a22003733i 4500 001 2429440
005 20250430100829.5
006 m o d
007 cr|nu||||||||
008 261011s2024 miu||||||m |||||||eng d
020 $a 9798346780014
035 $a (MiAaPQD)AAI31777867
035 $a AAI31777867
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Zheng, Huan, $e author. $3 3817288
245 1 0 $a Self-supervised Deep Learning for Image Denoising and Beyond / $c Huan Zheng.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 1 electronic resource (116 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-06, Section: B.
500 $a Advisors: Ji, Hui.
502 $b Ph.D. $c National University of Singapore (Singapore) $d 2024.
520 $a Images are central to the field of computer vision. However, their quality can be significantly compromised by the inevitable degradations introduced during capturing and transmission, leading to inaccuracies in subsequent analyses. Noise is one key issue existing in most degradations. This has motivated the development of robust and efficient denoising techniques. Of late, supervised deep learning methods, which learn a mapping from a noisy image to its clean version using extensive external datasets, have become prominent. However, these methods may perform poorly due to disparities between the training and testing datasets. Moreover, collecting a sizeable training datasets can be challenging under certain circumstances.Recently, there is increasing interest in unsupervised learning-based methods which relax the need for a paired noisy-clean dataset. In this thesis, we initially revisit the innovative Noise2Noise work, which trains a neural network (NN) with only paired independent noisy images. However, collecting such paired datasets still remain a daunting task. Motivated by this, we propose a novel unsupervised approach, Recorrupted to Recorrupted (R2R), which requires only unpaired noisy images for image denoising tasks. Assuming Gaussian noise, we proved that our R2R training is statistically equivalent to its supervised counterpart. Nevertheless, such Gaussian noise assumption often fails in real-world scenarios and we further extended our R2R for general noise distributions, culminating in the Extended R2R (ER2R) method. We validated ER2R's efficacy on video denoising tasks, using a specific NN architecture to efficiently and effectively exploit temporal redundancy. Notably, ER2R can be trained using only a single test video, thus negating the need for external training datasets. This approach meets the same criteria as traditional methods, but with significantly improved performance. Furthermore, we extended our ER2R method to inverse problems, specifically focusing on the non-blind deconvolution problem, and proposed a kernel-perturbation-based inference scheme for managing kernel errors in real-world scenarios. In conclusion, we propose an unsupervised denoising framework that is applicable to many problems including but not limited to image(video) denoising and linear image inverse problems. Compared to supervised learning and other unsupervised learning methods, our methods either offer superior performance or eliminate the need for paired training datasets while still yielding comparable results.
546 $a English
590 $a School code: 1883
650 4 $a Tomography. $3 836553
650 4 $a Deep learning. $3 3554982
650 4 $a Inverse problems. $3 3686950
650 4 $a Visualization. $3 586179
650 4 $a Normal distribution. $3 3561025
650 4 $a Microscopy. $3 540544
650 4 $a Mathematics. $3 515831
650 4 $a Medical imaging. $3 3172799
690 $a 0800
690 $a 0405
690 $a 0574
710 2 $a National University of Singapore (Singapore). $e degree granting institution. $3 3765631
720 1 $a Ji, Hui $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-06B.
790 $a 1883
791 $a Ph.D.
792 $a 2024
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31777867