Learning for feature selection and shape detection

The paper proposes a general framework for shape detection based on supervised symbolic learning. Differently from other visual systems exploiting machine learning, the proposed architecture does not follow the object segmentation - feature extraction and (learning based) classification approach. Instead, an initial data-driven processing selects points of interest in the scene by means of complex features which hypothesize the presence of the target shape; hypotheses are validated by a classifier defined by a machine learning algorithm. Learning is exploited not only for defining the model, i.e. the description of the target for the classifier, but also for defining the description language, i.e. the feature set useful in generating reliable object hypotheses. The proposed architecture of visual system has been implemented for an industrial application of unstructured shape detection: examples and results are reported in the paper.