Learning to classify seismic images with deep optimum-path forest

Abstract

Due to the lack of labeled information, clustering techniques have been paramount in the last years once more. In this paper, inspired by the deep learning phenomenon, we presented a multi-scale approach to obtain more refined cluster representations of the Optimum-Path Forest (OPF) classifier, which has obtained promising results in a number of works in the literature. Here, we propose to fill a gap in OPF-based works by using a deep-driven representation of the feature space. Additionally, we validated the work in the context of high resolution seismic images aiming at petroleum exploration, as well as in general-purpose applications. Quantitative and qualitative analysis are conducted in order to assess the robustness of the proposed approach.