Computer Vision, winter 2012-13 CS Department, Technion

Preview:

Citation preview

Image Pyramids

Computer Vision, winter 2012-13CS Department, Technion

TopicsThe Gaussian PyramidThe Laplacian PyramidApplications:

Pattern MatchingCoding (Compression)EnhancementBlending

Gaussian Pyramid

Gaussian PyramidApproximation filter is not Gaussian

A 5 tap binomial kernel:

2D extension:

1[1,4,6,4,1]

16

1 4 6 4 1

4 16 24 16 41

6 24 36 24 6256

4 16 24 16 4

1 4 6 4 1

Pattern Matching

Laplacian Pyramid

Laplacian PyramidApproximation of the Laplacian of each scaleConstructed by Difference of Gaussian (DoG)

layers

Laplacian Pyramid

Image Coding

Enhancement

Enhancement

Pyramid Blending

Step I - Create Blend Masks

Step II

• Weighted sum over each Laplacian pyramid scale (of the images) independently

• Gaussian pyramid of masks, or simple resizing of masks to correct dimensions

Blend 1Blend 1 MaskBlend 2Blend 2 Mask

Step IIIStandard Laplacian pyramid

reconstruction of blended scales

Implementation IssuesImage registrationFeatheringOver/under flows outside [0,1]MATLAB function: impyramid()

Only Gaussian pyramid‘reduce’ or ‘expand’ modes

MATLAB function: imresize()Useful for masks resizing

Recommended