Abstract
In this paper we describe a simple hybrid architecture of multi-model neural network aimed at enhancing the accuracy of classification in image interpretation problems. We adopt a modular architecture with one neural network dedicated to each class of the problem domain, allowing each of these neural modules to be built according to a different paradigm. The selection of the paradigm for each class is based on a benchmark among a set of competitor neural network models. We demonstrate experimentally the effectiveness of this approach in the problem of deforestation monitoring in the Amazon region.