Image Segmentation with Cascaded Hierarchical Models and Logistic Disjunctive Normal Networks

摘要: Contextual information plays an important role in solving vision problems such as image segmentation. However, extracting contextual and using it effective way remains a difficult problem. To address this challenge, we propose multi-resolution framework, called cascaded hierarchical model (CHM), which learns framework for At each level of the hierarchy, classifier is trained based on down sampled input images outputs previous levels. Our then incorporates resulting into to segment at original resolution. We repeat procedure by cascading improve segmentation accuracy. Multiple classifiers are learned CHM, therefore, fast accurate required make training tractable. The also needs be robust against over fitting due large number parameters during training. introduce novel classification scheme, logistic disjunctive normal networks (LDNN), consists one adaptive layer feature detectors implemented sigmoid functions followed two fixed layers logical units that compute conjunctions disjunctions, respectively. demonstrate LDNN outperforms state-of-the-art can used CHM object performance.