Image Feature Extraction: Global and Local Features, Haar Wavelets, SIFT, Study notes of Electrical and Electronics Engineering

Download Image Feature Extraction: Global and Local Features, Haar Wavelets, SIFT and more Study notes Electrical and Electronics Engineering in PDF only on Docsity!

Image Features

ECE 547 Lecture ECE 547 Lecture November 20, 2008 November 20, 2008 Mert Dikmen Mert Dikmen Image Formation and Processing Group Image Formation and Processing Group

Overview

Definition Definition Global Features Global Features Local Features Local Features Dense Local Features Dense Local Features Haar Wavelets Haar Wavelets Sparse Local Features Sparse Local Features SIFT SIFT

Example 1

Salmon or Seabass? Salmon or Seabass? Images courtesy of Duda, Heart & Stork, ”Pattern Classification”, Wiley Interscience

Salmon vs Seabass cont...

Salmon vs Seabass cont...

What are features?

low- to mid-level abstractions of the information low- to mid-level abstractions of the information contained in the image contained in the image Designing a feature mapping: Designing a feature mapping: ● Perceptual intuitionPerceptual intuition ● Prior knowledgePrior knowledge ● Viewpoint invarianceViewpoint invariance ● Driven by the goal of the problemDriven by the goal of the problem

Global Features:

1. Mean, other Moments & Variance

n

−∞ ∞

x

n

dF  x = E  x

n

n

NM

i , j

I  i , j 

n Moments in statistics: Moments in statistics: Image Moments:Image Moments:

Global Features:

2. Image Histogram

Hist  k =

∑ x , y

 I  x , y  , k 

k ∈{0,1, , 255 }

Dense Local Features

● Computed from a predefined grid of local patchesComputed from a predefined grid of local patches ● Convey information on local structureConvey information on local structure ● Good for detecting rigid objectsGood for detecting rigid objects ● Trade off locality vs informativenessTrade off locality vs informativeness ● Usually high dimensionalUsually high dimensional ● Even more than the number of pixelsEven more than the number of pixels ● Very costly to computeVery costly to compute

Dense Local Features:

1. Edges

Dense Local Features:

1. Filtering

G  x , y = e − x 2  y 2  ∂ ∂ x G  x , y =−2x e − x 2  y 2 

Dense Local Features:

Filtering

∂ ∂ x G  x , y =−2x e − x 2  y 2  = G 0 ᐤ ∂ ∂ y G  x , y =−2y e − x 2  y 2  = G 90 ᐤ G  =cos  G 0 sin  G 90

Dense Local Features:

2. Filtering

(^44) thth order filtersorder filters (^55) thth order filtersorder filters (^11) stst^ order filtersorder filters

Dense Local Features:

2. Filtering