fence removal from images

Upload: jerrykorulla91

Post on 01-Jun-2018

217 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/9/2019 Fence Removal From Images

    1/20

  • 8/9/2019 Fence Removal From Images

    2/20

    What is De-Fencing?

    Fences: near-regular foreground patterns that are often unwanted, but unavoidabimagery.

  • 8/9/2019 Fence Removal From Images

    3/20

    What is De-Fencing?De-fencing: Process of automatically removing ‘fences’ from an image.

  • 8/9/2019 Fence Removal From Images

    4/20

    ApplicationsAnimal photography in zoo.

    Sports photography

    Photography of buildings behind fences

  • 8/9/2019 Fence Removal From Images

    5/20

    Present Techniques1. "Image De-fencing“, Yanxi Liu, Tamara Belkina, James H. Hays, and Roberto Lub

    Computer Vision and Pattern Recognition Conference (CVPR) 2008.

    2. “Image De-fencing Revisited ”, Minwoo Park, Kyle Brocklehurst, Robert T. CollinsLiu , Asian Conference on Computer Vision (ACCV) 2010.

  • 8/9/2019 Fence Removal From Images

    6/20

    Steps in De-fencing1. Finding a lattice.

    2. Classifying pixels as foreground or background.

    3. Filling the background holes with texture inpainting.

  • 8/9/2019 Fence Removal From Images

    7/20

    Finding the Lattice1. Assign neighbor relationships among a set of interest points.

    2. Use the strongest cluster of repeated elements to propose new, visually similar points.

    3. Higher-order constrains are used to promote geometric consistency between paassignments.

    Online learning using a support vector machine can be performed to improve the

    of lattice points and for foreground segmentation.

  • 8/9/2019 Fence Removal From Images

    8/20

    Foreground/background Separation

    a) Align the texels

    b) Compute the standard deviation o

    this stack of texels.

    c) The cluster which has the lowest v

    taken as foreground and the rest b

    d) Create a mask containing the foreg

  • 8/9/2019 Fence Removal From Images

    9/20

    Texture InpaintingΩ= Target Region   Φ= Source Region   δΩ= Cont

    1. The square template Ψp ϵ Ω centred at the point p is to be

    2. The best-match sample from the source region comes fromΨqϵΦ which is most similar to those parts that are already

    3. To propagate the isophote inwards, transfer the pattern fromatch source patch to Ψp.

    Additional techniques: Multi-view inpainting

    Symmetry augmented inpainting

  • 8/9/2019 Fence Removal From Images

    10/20

    A Novel Spatial Domain Approach foFence Detection.

    1. Find the edges of the input image.

    2. Obtain Hough Transform of the edges.

    3. Detect the peaks in the Hough transform.

    4. Find the dominant angles in the edge map from the peaks in the Hough transfor

    5. Separate the fences from the image making use of the information about the doangles of the fence.

  • 8/9/2019 Fence Removal From Images

    11/20

    Input Image

  • 8/9/2019 Fence Removal From Images

    12/20

    Edges

  • 8/9/2019 Fence Removal From Images

    13/20

    Hough Transform of Edges

  • 8/9/2019 Fence Removal From Images

    14/20

  • 8/9/2019 Fence Removal From Images

    15/20

    Strength of edges in different directi

  • 8/9/2019 Fence Removal From Images

    16/20

  • 8/9/2019 Fence Removal From Images

    17/20

    Work Done

    Image Inpainting

    Exemplar based inpainting

    Fence Detection

    Mean Shift belief propagation

    Morphological operations Frequency domain technique

    Hough Transform

  • 8/9/2019 Fence Removal From Images

    18/20

    Further Work

    Refine the developed spatial domain technique to make the detection more accu

    Develop other techniques for fence detection.

  • 8/9/2019 Fence Removal From Images

    19/20

    Reference1. "Image De-fencing“, Yanxi Liu, Tamara Belkina, James H. Hays, and Roberto Lub

    Computer Vision and Pattern Recognition Conference (CVPR) 2008.

    2. “Image De-fencing Revisited ”, Minwoo Park, Kyle Brocklehurst, Robert T. CollinsLiu , Asian Conference on Computer Vision (ACCV) 2010.

    3. “Discovering texture regularity as a higher-order correspondence problem” , J. Leordeanu, A. Efros, and Y. Liu, European Conference on Computer Vision (ECCV’

    4. ‘Object Removal by Exemplar -based Inpainting’, Antonio Criminisi, Patrick PereKentaro Toyama, IEEE Computer Vision and Pattern Recognition (CVPR), June 20

  • 8/9/2019 Fence Removal From Images

    20/20