Created by W.Langdon from gp-bibliography.bib Revision:1.5628
Most texture feature extraction methods are derived from human intuition after much contemplation. Texture feature extraction remains a challenging problem due to the diversity and complexity of natural textures. In this thesis we investigate the evolution of feature extraction programs using tree based genetic programming.
Our main hypothesis is that given the right fitness evaluation, it may be possible to generate new feature extraction programs independent of human intuition from basic properties of images such as pixel intensities, histograms and pixel positions. We used tree based genetic programming and a learning set of thirteen Brodatz textures to evolve feature extraction programs. We have investigated three kinds of inputs/terminals: raw pixels, histograms and a spatial encoding. The function set consisted of +,- to facilitate the analysis of the evolved programs. Fitness is computed with a novel application of clustering. A program in the population is applied to a selection of images of two textures in the learning set. If the program delivers widely separated clusters for the two textures, it is considered to be very fit.",
The evolved programs were evaluated by classification accuracy on the testing sets. Raw pixel input gave a classification accuracy of 50percent for task 1 and 45percent for task 3. Histogram input gave a classification accuracy of 81percent and 75percent for these tasks while the spatial encoding gave accuracies of 75percent and 61percent. The histogram representation was found to be the most effective representation. The evolved programs were compared with 18 human derived methods on tasks 1 and 3. The accuracy of the evolved programs was ranked 14 out of 19 for task 1 and 9 out of 19 for task 3. Task 2 was only performed using histogram inputs and the accuracy was 100percent compared with 95percent for the grey level co-occurrence method. These results indicate that, on these tasks, the evolved feature extraction programs are competitive with human derived methods.
Task 4, malt classification, is a difficult real world problem. We used the best performing input, histograms, for this task. We obtained a classification accuracy of 67percent which is better than the Gabor and Haar methods but worse than the gray level co-occurrence matrix and the grey level run length methods. However, when we combined the evolved features with human derived features, we improved the classification accuracy by 15percent. This suggests that the evolved features have captured texture regularities not captured in the human derived methods. The contribution of the evolved features towards the improved accuracy was confirmed when the combined evolved and human derived feature set was subjected to feature selection. There was a high percentage of evolved features among the selected features.
The value of our approach lies in the fact that feature extraction programs can be evolved from simple inputs such as histograms and arithmetic operations without much domain knowledge. From a practitioner point of view, our set of programs has the advantage of not requiring the user to set the parameter values as required by many human derived methods. For researchers, our approach shows that it is possible to evolve, from simple inputs, feature extraction programs that can perform as well as those derived by human intuition.",
Genetic Programming entries for Brian Lam