This preview shows pages 1–12. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: 1 Perceptual and Sensory Augmented Computing Lecture Layout Image features Edges Junctions & Corners Blobs Ridges Image descriptors SIFT features Goals for a lowlevel image representation • Compact Reduce the number of processed image locations 000000011111100000111111111000 = 000000 01 1111 10 000 01 1111111 10 00 = 7x0 5x1 5x0 7x1 3x0 • Generic • Sufficient No need to look back into the image 2 2 • Boundaries: 1D structures in 2D from O(N^2) to O(N) • Corners/Junctions: 0D structures: O(1) (`Boundaries of boundaries’) Boundaries 3 Attneave, 1954 A biref anlaogy wtih txet • Waht mttares is waht hpapens on wrod baounrdies • Mocpera iwht htsi (compare with this) • Massive evidence of our visual system employing boundary detection 4 3 Symmetries 5 Perceptual and Sensory Augmented Computing Lecture Layout Image features Edges Junctions & Corners Blobs Ridges Image descriptors SIFT features 4 Edges in images • Edges are not only object boundaries… … and object boundaries do not always cause edges 7 surface normal discontinuity shape depth discontinuity boundaries surface color discontinuity texture illumination discontinuity shadows Object Boundaries vs. Edges Background Texture Shadows 5 Perceptual and Sensory Augmented Computing Lecture Layout Image features Edges 1D edge detection 2D edge detection LaplacianofGaussian Canny edge detection Recent research Junctions & Corners Blobs Ridges Image descriptors SIFT features • Edges: discontinuities in the image intensity function • Estimate derivative, find its peaks and we are done Finding edges image 6 Finding edges in real images • Consider images corrupted with Gaussian noise image Why does this happen? • Consider Fourier transform Magnitude, F Add noise Multiply FTs 7 Last lecture: smoothing to remove noise 13 • First convolve signal with Gaussian filter d/dx Convolve Last lecture: smoothing to remove noise 14 • Large scales: increased robustness to noise • But poor localization accuracy Convolve d/dx 8 • Convolution with Gaussian: windowing of FT Frequencydomain explanation 15 Multiply Multiply Smaller kernel width: higher frequencies • Scaling property of Fourier transform: 16 Multiply Multiply 9 • Find edges at minima/maxima of derivative Condition for extremum: derivative equal to zero Edge localization 17 Convolve d/dx d/dx • Differentiation, smoothing: linear operators (filters) Cascade of filters can be replaced with a single filter • Differentiation property of convolution: • Convolve with and detect zerocrossings Allinone filter 18 10 19 1D edge detection: summary 1st derivative 2nd derivative Multiscale 1D edge detection • As σ increases, the smoothing increases and only the central edge survives. 11 Perceptual and Sensory Augmented Computing Lecture Layout Image features Edges 1D edge detection...
View
Full
Document
This note was uploaded on 11/06/2010 for the course CSE 527 taught by Professor Ab during the Fall '09 term at Cornell.
 Fall '09
 Ab

Click to edit the document details