Please find attached our entry for the competition.
 
Regards,
Brian Lam

 
Brian Lam
RMIT University
GPO Box 2476V
Melbourne, VIC 3001
Australia
blam@cs.rmit.edu.au
tel. 613-9925-3781
 
Vic Ciesielski
RMIT University
GPO Box 2476V
Melbourne, VIC 3001
Australia
vc@cs.rmit.edu.au
tel. 613-9925-2926
 
"Discovery of Human-Competitive Image Texture Feature Programs Using Genetic programming"
 
Abstract:
In this paper we show how genetic programming can be used to discover useful texture feature extraction algorithms. Grey level histograms of different textures are used as inputs to the evolved programs. One dimensional K-means clustering is applied to the outputs and the tightness of the clusters is used as the fitness measure. To test generality, textures from the Brodatz library were used in learning phase and the evolved features were used on classification problems based on the Vistex library. Using the evolved features gave a test accuracy of 74.8% while using Haralick features, the most commonly used method in texture classification, gave an accuracy of 75.5% on the same problem. Thus, the evolved features are competitive with those derived by human intuition and analysis. Furthermore, when the evolved features are combined with the Haralick features the accuracy increases to 83.2%, indicating that the evolved features are finding texture regularities not used in the Haralick approach.

Criteria for Entry:

We believe our results meet criteria D and F. 

For criteria D, we believe the feature extraction method evolved by GP is a novel method which is publishable in the computer vision literature as a new scientific result independent of the fact that the result was mechanically generated.

For criteria F, we believe that our result is superior to many well  known methods in the field.  These methods were considered achievements at the time they were discovered.

The problem of generating texture features that can be used for accurate, generalised visual texture classification has been addressed by many mathematicians and computer scientists since the early 1970s.  The first major advance was by Haralick who proposed a set of features based on the grey level co-occurrence matrix, a data structure that he invented.  Haralick  features are still widely used today and are generally used as a benchmark for new approaches.  There have been a number of attempts to improve on the Haralick approach and to find completely new alternatives.  In a major study, Wagner compared all of the major approaches to texture classification on the same set of problems.  We have applied our evolved feature extraction programs to this set of problems.  As can be seen from the table below our results are better than 7 methods, very similar to 6 methods and not quite as good as 6 methods.

Local features      47%
Fractal 2           48%
Fractal 1           53%
Sun & Wee           59%
Amadasun            67%
Markov               70%
Galloway            70%
Mao & Jain          72%
Pikaz & Averbuch    73%
Dapeng              73%
GP Evolved Features 73%   <--------Our accuracy
Gabor               73%
Laine               73%
Haralick            79%
Laws                80%
Fourier coeff       80%
Unser               81%
Amelung             82%
Chen                84%