본문 바로가기
HOME> 보고서 > 보고서 검색상세

보고서 상세정보

Development of an algorithm for salient region detection using deep learning
Development of an algorithm for salient region detection using deep learning

  • 과제명

    딥러닝을 이용한 핵심 영역 검출 알고리즘 개발

  • 주관연구기관

    한국과학기술원
    Korea Advanced Institute of Science and Technology

  • 보고서유형

    최종보고서

  • 발행국가

    대한민국

  • 언어

    대한민국

  • 발행년월

    2015-12

  • 과제시작년도

    2015

  • 주관부처

    미래창조과학부
    KA

  • 사업 관리 기관

    한국과학기술원
    Korea Advanced Institute of Science and Technology

  • 등록번호

    TRKO201600002326

  • 과제고유번호

    1711032181

  • DB 구축일자

    2016-06-04

  • DOI

    https://doi.org/10.23000/TRKO201600002326

  • 초록 


    1. Research Goals
    A. Research Background
    Human visual system unconsciously finds a distinctive part in a scene to focus on ...

    1. Research Goals
    A. Research Background
    Human visual system unconsciously finds a distinctive part in a scene to focus on for a complex process such as segmentation and recognition. Similarly, in computer vision, salient region detection tries to automatically find an informative and distinctive region in an image based on some criteria as shown in Figure 1.
    Saliency detection aims to detect distinctive regions in an image that draw human attention. This topic has received a great deal of attention in computer vision and cognitive science because of its wide range of applications such as content-aware image cropping [21] and resizing [3], video summarization [23],object detection [19], and person re-identification [30]. Various papers such as DRFI [12], GMR [29], DSR [16], RBD [31], HDCT [14], HS [28] and GC [7] utilize low level features such as color, texture and location information to investigate characteristics of salient regions including abjectness, boundary convexity, spatial distribution, and global contrast.
    Recent trends in salient region detection utilize learning based approaches, which were first introduced by Liu et al. [18]. Liu et al. were also the first group to released a benchmark dataset (MSRA10K) with ground truth evaluation.
    Following this work, several representative benchmarks with ground truth evaluation were released. These benchmarks include ECSSD [28], Judd [13],THUR15K [6], DUTOMRON[29], PASCAL-S [17], and FT [1]. They cover rich variety of images containing different scenes and subjects. In addition, each one exhibits different characteristics. For example, the ground truth of the MSRA10K dataset are binary mask images which were manually segmented by human, while the ground truth of the FT [1] dataset were determined by human fixation.
    Discriminative Regional Feature Integration(DRFI) [12], Robust Background Detection(RBD) [31], Dense and Sparse Reconstruction(DSR) [16], Markov Chain(MC) [11], High Dimensional Color Transform(HDCT) [14], and Hierarchical Saliency(HS) [28] are the top 6 models for salient region detection reported in the benchmark paper [5]. These algorithms consider various heuristic priors such as the global contrast prior [28] and the boundary prior [12] and often generate high-dimensional features to increase discriminative power [14, 12] to distinguish salient regions from non-salient regions. These methods are all based on handcrafted low level features without deep learning.
    Deep learning has emerged in the field of saliency detection last year. Several methods that utilize deep learnings for saliency detection were simultaneously proposed. This includes Multiscale Deep Feature(MDF) [15], MultiContext Deep Learning(MCDL) [20], and Local Estimation and Global Search(LEGS) [26]. They utilized high level features from the deep convolutional neural network (CNN) and demonstrated superior results over previous works that utilized only low level features. MDF and MCDL utilize superpixel algorithms, and query each region individually to assign
    saliency to superpixels. For each query region, MDF generates three input images that cover different scopes of an input image, and MCDL uses sliding windows with deep CNN to compute the deep features of the center superpixel. LEGS first generates an initial rough saliency mask from deep CNN and refines the saliency map using an object proposal algorithm.


    ...


  • 목차(Contents) 

    1. COVER ... 1
    2. Part Ⅰ. General Information ... 2
    3. CONTENTS ... 3
    4. 1.Research Goals ... 4
    5. A.Research Background ... 4
    6. B.Research Goals ... 6
    7. 2.Research Desc...
    1. COVER ... 1
    2. Part Ⅰ. General Information ... 2
    3. CONTENTS ... 3
    4. 1.Research Goals ... 4
    5. A.Research Background ... 4
    6. B.Research Goals ... 6
    7. 2.Research Description ... 7
    8. A.Research goal ... 7
    9. B.Research progress ... 8
    10. C.Approach method ... 9
    11. D.Research Outcomes ... 16
    12. 3.Research Achievement ... 17
    13. A.Research Performance ... 17
    14. B.Expected Research Outcome ... 22
    15. 4.Reference ... 23
    16. End of Page ... 25
  • 참고문헌

    1. 전체(0)
    2. 논문(0)
    3. 특허(0)
    4. 보고서(0)

 활용도 분석

  • 상세보기

    amChart 영역
  • 원문보기

    amChart 영역