Integração de operadores conexos baseados em árvores de componentes em redes neurais supervisionadas

Abstract

Computer vision is a subfield of artificial intelligence aimed at enabling computers to automatically interpret and understand images and videos. It plays a key role in various applications, ranging from medical diagnostics and autonomous vehicles to security sys- tems and precision agriculture. In this context, deep neural networks stand out for their generalization power in computer vision tasks, although they present limitations in terms of interpretability. On the other hand, mathematical morphology offers structural tools for image analysis and segmentation based on the connectivity and shape of objects, thus providing greater interpretability. This dissertation proposes an integration of the genera- lization capabilities of neural networks with the interpretability of connected filters, which are filtering tools that operate through the fusion of regions in an image. Recent studies highlight challenges in combining these techniques, especially due to the difficulty in opti- mizing the parameters used by connected filters, which are traditionally based on binary functions. Nevertheless, some approaches have been proposed to overcome this limitation, paving the way for further research and enabling this work to explore such integration. The proposed approach relies on a neural network architecture with a layer that applies a morphological filter whose parameters are incorporated into the model itself, making it possible to optimize them through machine learning techniques. Experiments focused on an image segmentation problem demonstrate that the method can produce effective segmentations, successfully balancing performance and interpretability.