Download Computer Vision Lecture Notes: Texture and Filter Banks - Prof. Pradeep Sen and more Assignments Electrical and Electronics Engineering in PDF only on Docsity!
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
ECE 516 / CS 532
Computer Vision
Pradeep Sen Advanced Graphics Lab
Class 10 February 25, 2009
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Announcements
HW 2 has been posted
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Last time
Edge detection algorithms
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Today
Texture
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
What is texture?
Organized patterns of regular subelements in an image that represent material properties
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Application of texture
Texture segmentation – breaking up an image into components with constant texture Texture synthesis – Construct large regions of texture from small sample images Shape from texture – Recovering surface orientation or shape from texture
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Representing textures
The regular subelements that compose a texture are called textons
So one way to represent a texture is to decompose it into the textons that make it up
However, there is no canonical set of textons that make up all textures ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Human perception of textons
Can you make out the different textures in this image?
From Bergen and Landy, “Computational Modeling of Visual Texture Segregation”
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Representing texture
An easier way is to represent the texture by projecting it onto a specific basis
This can be done with a filter bank , a series of different filters that are applied to the image
Convolving a filter with the image is equivalent of applying
One common filter bank consists of spots and bars
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Spots and bars filter bank
A spot, given by a weighted sum of three concentric symmetric Gaussians, with weights 1, -2, 1 and sigma 0.62, 1, 1. A second spot, given by a weighted sum of two concentric, symmetric Gaussians, with weights 1 and -1 and sigma 0.71, 1. A set of six series of oriented bars made up of the weighted sum of three oriented Gaussians
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Spot and bars filter bank
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Example
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Laplacian pyramid
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Laplacian pyramid
A bit of a misnomer – no differential operators are used Each layer of the Laplacian pyramid can be thought of as the response of a band-pass filter
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Frequency response of Laplacian pyramid
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Localized spatial filters
The problem with Fourier transforms is that the basis functions span the entire image
One solution is to use locally responsive filters such as Gabor filters
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Gabor filters
Essentially Fourier basis multiplied by a Gaussian to limit their influence
They come in pairs, called quadrature pairs, where one recovers symmetric components, the other asymmetric components
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Gabor filters
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Oriented pyramids
Constructed by applying oriented filters to layers of the Laplacian pyramid
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Oriented pyramids
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Frequency response of oriented pyramid
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Texture synthesis
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Texture synthesis
Efros and Leung (1999) – Local probabilistic models
For each pixel to synthesize, look at its neighborhood.
Find comparable neighborhoods in the sample image
Draw from those pixel samples through a random distribution
ECE516 / CS532 Computer Vision Pradeep Sen Class 10 – February 25, 2009
Reading
Forsyth, Ch 9
