We aim at synthesizing the complete image analysis and interpretation procedures without splitting it explicitly into stages of feature extraction and classification. As a result, the learning process is no more limited to the image representation and features predefined by the human expert, but encompasses also the image processing and analysis. Thus, we follow here the paradigm of direct approach to pattern recognition, occupied so far mostly by the artificial neural networks. However, instead of subsymbolic reasoning and gradient-based learning, we employ the paradigm of evolutionary computation for the search of the space of image analysis programs.

From the machine learning viewpoint, we focus on the paradigm of supervised learning from examples, employing the genetic programming for the hypothesis representation and for the search of hypothesis space. The candidate procedures performing image analysis and recognition are evaluated on a set of training cases (images). In the talk, we will also discuss some related topics (non-scalar evaluation and local optimization) and present preliminary experimental results obtained on real-world tasks.