DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Yoo, Chang D. | - |
dc.contributor.advisor | 유창동 | - |
dc.contributor.author | Lee, Seungjae | - |
dc.date.accessioned | 2021-05-12T19:40:42Z | - |
dc.date.available | 2021-05-12T19:40:42Z | - |
dc.date.issued | 2020 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=909414&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/284185 | - |
dc.description | 학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2020.2,[v, 56 p. :] | - |
dc.description.abstract | Humans search objects through collaboration with sensory organs and the brain and understand the environment. For computers to search and understand their surroundings, we need to define search targets and provide searchable structures by extracting and processing features of them. For this objective, visual search has developed and studied feature design and extraction. However, recent developments in deep learning have changed the visual search from feature design to network architecture design. In this dissertation, we examine the recent shift from feature engineering to architecture engineering for visual search and propose an efficient feature extraction for visual place recognition, an energy-efficient architecture for object detection and neural architecture search for visual search. First, we propose visual place recognition based on feature engineering. The visual place recognition is a technology to localize the position by using the visual feature. In this study, the whole process of visual place recognition from reference images collection to application has been designed and implemented in Meyong-dong and Garosu-gil. For efficient visual place recognition, informative reference image selection, feature selection and partial gradient computation are proposed, and the implemented system has shown the possibility of the proposed method. Second, we address an efficient network design by reviewing ImageNet Large Scale Visual Recognition Challenge (ILSVRC). The ILSVRC had led visual search and shifted visual search from feature engineering to architecture engineering. But the architecture engineering needs a huge amount of resources and efforts. Moreover, related architectures in the challenge are difficult to support mobile applications. To overcome this limitation, an efficient architectures have been proposed and studied. We propose energy-efficient object detection architecture and its effectiveness has been verified in Low Power ImageNet Recognition Challenge (LPIRC). Third, Neural Architecture Search (NAS) has been studied. NAS designs and searches the architecture with machine learning automatically, but it sometimes requires an extensive amount of computing power. To overcome this problem, an efficient NAS has been getting more attention and a differentiable method such as Differentiable Architecture search (DARTS) has improved the search efficiency. In this study, we examine the reproducibility of DARTS by analyzing random initialization, architecture selection and optimization in the training process of DARTS. To improve DARTS, Topology-based Progressive Neural Architecture Search (TP-NAS) is proposed. TP-NAS first determines the connection of the search space and search the operation progressively for the architecture. The proposed method is verified by the experiments on CIFAR-10 and it shows the transferability on ImageNet and CIFAR-100. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | visual search▼afeature extraction▼aneural architecture design▼aneural architecture search▼adeep learning | - |
dc.subject | 비주얼 검색▼a특징 추출▼a네트워크 설계▼a네트워크 탐색▼a딥러닝 | - |
dc.title | Study on efficient feature extraction and neural architecture engineering for visual search | - |
dc.title.alternative | 시각적 검색을 위한 효율적인 특징 및 구조 설계 방법 연구 | - |
dc.type | Thesis(Ph.D) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :전기및전자공학부, | - |
dc.contributor.alternativeauthor | 이승재 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.