Neural networks in image pyramids
Horst Bischof, Axel J. Pinz · 2003
The authors present a novel neural network model for visual information processing. The model uses a hierarchical network with local connectivity as a stem network. This network generates hypotheses about the expected image content, and then selectively uses small neural network modules on parts of the image to check these hypotheses. The resulting neural network is able to use different spatial resolutions, and is both modular and hierarchical. Applying this model to remotely sensed image classification (Landsat TM) is described. A slightly better classification accuracy was achieved at reduced computational cost, compared to classification without the model.>