DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Kim, Changick | - |
dc.contributor.advisor | 김창익 | - |
dc.contributor.author | Ko, Kangwook | - |
dc.date.accessioned | 2023-06-26T19:33:52Z | - |
dc.date.available | 2023-06-26T19:33:52Z | - |
dc.date.issued | 2023 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=1032865&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/309874 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2023.2,[iii, 20 p. :] | - |
dc.description.abstract | Image deblurring is a classic computer vision task which aims to recover the latent sharp image from blurry input. A number of deep-learning-based methods have been introduced to solve this problem, using multi-scale, multi-patch, or multi-temporal approaches. These approaches achieved good performances in image deblurring. However, when we focus on inference speed, these approaches have some limitations. In this paper, we propose a fast and accurate single-stage network named FDANet. FDANet can handle dynamic blurs in computationally efficient operations by exploiting frequency-domain information with frequency-domain attention. The proposed frequency-domain attention block separately attends regions with different degrees of blurs. The shallowest version of our network has the shortest inference time among existing deblurring methods, and the deeper counterpart surpasses other methods with comparable processing time. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | Image deblurring▼aFrequency-domain attention▼aComputer vision | - |
dc.subject | 영상 디블러링▼a주파수 영역 어텐션▼a컴퓨터 비전 | - |
dc.title | Frequency-domain attention network for fast and accurate image deblurring | - |
dc.title.alternative | 고속 영상 디블러링을 위한 주파수 영역 어텐션 신경망 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :전기및전자공학부, | - |
dc.contributor.alternativeauthor | 고강욱 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.