Page 136 - 49A Field Guide to Genetic Programming

122 12 Applications

signals (Sharman and Esparcia-Alcazar, 1993). Local search (simulated an-
nealing or gradient descent) can be used to adjust or fine-tune “constant”
values within the structure created by genetic search (Smart and Zhang,
2004).
Yu and Bhanu (2006) have used GP to preprocess images, particularly
of human faces, to find regions of interest for subsequent analysis. See also
(Trujillo and Olague, 2006a).
Zhang has been particularly active at evolving programs with GP to
visually classify objects (typically coins) (Zhang and Smart, 2006). He has
also applied GP to human speech (Xie, Zhang, and Andreae, 2006).
“Parisian GP” is a system in which the image processing task is split
across a swarm of evolving agents (“flies”). In (Louchet, 2001; Louchet,
Guyon, Lesot, and Boumaza, 2002) the flies reconstruct three dimensions
from pairs of stereo images. For example, in (Louchet, 2001), as the flies
buzz around in three dimensions their position is projected onto the left and
right of a pair of stereo images. The fitness function tries to minimise the
discrepancy between the two images, thus encouraging the flies to settle on
visible surfaces in the 3-D space. So, the true 3-D space is inferred from
pairs of 2-D images taken from slightly different positions.
While the likes of Google have effectively indexed the written word, for
speech and pictures indexing has been much less effective. One area where
GP might be applied is in the automatic indexing of images. Some initial
steps in this direction are given in (Theiler, Harvey, Brumby, Szymanski,
Alferink, Perkins, Porter, and Bloch, 1999).
To some extent, extracting text from images (OCR) can be done fairly
reliably, and the accuracy rate on well formed letters and digits is close
to 100%. However, many interesting cases remain (Cilibrasi and Vitanyi,
2005) such as Arabic (Klassen and Heywood, 2002) and oriental languages,
handwriting (De Stefano, Cioppa, and Marcelli, 2002; Gagne and Parizeau,
2006; Krawiec, 2004; Teredesai and Govindaraju, 2005) (such as the MNIST
examples), other texts (Rivero, nal, Dorado, and Pazos, 2004) and musical
scores (Quintana, Poli, and Claridge, 2006).
The scope for applications of GP to image and signal processing is almost
unbounded. A promising area is medical imaging (Poli, 1996b). GP image
techniques can also be used with sonar signals (Martin, 2006). Off-line work
on images includes security and verification. For example, Usman, Khan,
Chamlawi, and Majid (2007) have used GP to detect image watermarks
which have been tampered with. Recent work by Zhang has incorporated
multi-objective fitness into GP image processing (Zhang and Rockett, 2006).
In 1999 Poli, Cagnoni and others founded the annual European Work-
shop on Evolutionary Computation in Image Analysis and Signal Processing
(EvoIASP). EvoIASP is held every year with the EuroGP. Whilst not solely
dedicated to GP, many GP applications have been presented at EvoIASP.