Feature Learning from Image Markers for Object Delineation

• Italos Estilon de Souza Unicamp
• Barbara C. Benato Unicamp
• Alexandre Xavier Falcão Unicamp

Resumo

Convolutional neural networks (CNNs) have been used in several computer vision applications. However, most well-succeeded models are usually pre-trained on large labeled datasets. The adaptation of such models to new applications (or datasets) with no label information might be an issue, calling for the construction of a suitable model from scratch. In this paper, we introduce an interactive method to estimate CNN filters from image markers with no need for backpropagation and pre-trained models. The method, named FLIM (feature learning from image markers), exploits the user knowledge about image regions that discriminate objects for marker selection. For a given CNN's architecture and user-drawn markers in an input image, FLIM can estimate the CNN filters by clustering marker pixels in a layer-by-layer fashion -- i.e., the filters of a current layer are estimated from the output of the previous one. We demonstrate the advantages of FLIM for object delineation over alternatives based on a state-of-the-art pre-trained model and the Lab color space. The results indicate the potential of the method towards the construction of explainable CNN models.

Palavras-chave: Object delineation, convolutional neural networks, feature extraction