Optimization of evolutionary constructed features for image recognition

We address the issue of image recognition from the position of performance and robustness. Lately, impressive advancements have been made in this field, largely due to the substantial increase in computational power available at the consumer price. However, there still remains a need to optimize classification algorithms while keeping them sensibly robust to various space-time variations. We propose an algorithm, inspired by the work of Lillywhite et al. [1]. We build our filter bank and select our regions of interest via the framework of genetic algorithms. Then we treat constructed filters as weak hypothesis classifiers and use Adaboost algorithm to generate a strong classifier. We show that the length of the chromosome which comprises a filter and a number of different image transformations used as a part of a filter can be substantially reduced while still yielding a sensible recognition result.