In this work, three variations of hierarchical topologies of Convolutional Neural Networks (CNNs), two of which being original proposals introduced by this work, were tested to assess their impact on image classification problems. The hierarchical structure groups the images based on the semantic meaning of the classes, from the coarsest classes to the finest classes, forming hierarchical levels. The hierarchical models made were compared to a baseline regular CNN on benchmark image classification datasets, the Fashion-MNIST and CIFAR-100 datasets. Another contribution of this work is a new training strategy for hierarchical CNNs, that aims to be simple to implement and to produce a smooth loss during training, increasing stability, while maintaining characteristics like the transitioning from coarse-to-fine level emphasis during training, learning first high-level details and then specific details that differentiate the fine level classes. The hierarchical models produce outputs for each hierarchical level, which can lead to more interpretable results. Results suggest that providing semantic hierarchies can improve fine level accuracy on CNNs, while bringing relevant hierarchical information from their other coarser level outputs.