Aprendizagem de filtros conexos por redes neurais usando árvores de componentes

Abstract

Connected filters are widely recognized for their ability to preserve contours in ima- ges. A common approach to implementing them uses image representations based on component trees, which allow the calculation of characteristic attributes of the connected components represented by the tree nodes. These attributes can be used to filter specific nodes based on thresholds and then reconstruct the filtered image. Despite their rele- vance, the literature contains few initiatives that directly integrate the machine learning of connected filters within the context of neural networks. This dissertation proposes an innovative approach to optimize filtering in component trees by integrating them direc- tly into the neural network learning process. Instead of the traditional boolean function used to select the nodes, the approach employs a continuous and parameterized func- tion, meeting the requirements for neural network training. The experiments conducted demonstrate that the proposed method effectively learns connected filters, showing consis- tent performance across different image datasets, attributes, and training configurations, thereby consolidating its applicability and efficiency.