[Main]  [Publications]  [Abstracts and Posters]  [Software]  [Awards and Scholarships]  [Computer Vision Related Notes]


Matching Pursuit

Matching pursuit [1,2] is a greedy sub-optimal strategy for selecting basis vectors from an overcomplete set of basis vectors (termed the dictionary) to compute an adaptive compact representation of the input signal. At each iteration the algorithm selects the dictionary element that provides the maximum projection of the residual image.  The example reconstructions below were generated using a dictionary consisting of a two-dimensional Gaussian (at 3 scales) and a second derivative of a Gaussian taken along one direction (at 3 scales and 8 orientations).

 

 

Example 1

 

Figure 1a: Original Image (couple_orig.jpg)

180 x 196 x 8 BPP

Figure 1b: Matching Pursuit Reconstruction

1300 basis vectors (32 x 32 kernels)

 

 

Example 2

 

Figure 2a: Original Image (050510rain_orig.jpg)

150 x 200 x 8 BPP

Figure 2b: Matching Pursuit Reconstruction

2750 basis vectors (32 x 32 kernels)

 


[1]  Mallat, S. and Zhang, Z., Matching pursuits with time-frequency dictionaries. IEEE Transactions on Signal Processing, 41(12), December 1993.

 

[2]  Bergeaud, F. and Mallat, S., Matching Pursuit of Images, SPIE, Orlando, 1995.


Last update: July 1, 2006.