Interactive Inference for Spatial Image Analysis

• Activity Number: 434
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Hannah Director* and James Gattiker
• Companies: Los Alamos National Laboratory and Los Alamos National Laboratory
• Keywords: Interactive Analysis ; Machine Learning ; Spatial Statistics ; Image Segmentation

Abstract

Statistical modeling has traditionally advocated making inferences based on a specified model, without the benefit of expert guidance. Interactive inference seeks to overcome this limitation by combining statistical techniques and the knowledge of experts in a principled manner. Through human-in-the-loop analyses, or iterative processes where an expert provides information to be incorporated into a statistical model, human feedback can be fully integrated into an analysis. In this talk, I will compare models for interactive image segmentation. Segmentation seeks to identify distinct regions in an image, such as the edges of individual crystals in a materials image. In this interactive approach, an expert labels a small number of pixels which are used along with features of the image to develop a spatial model of the image. Uncertain areas of the segmentation are identified and additional labels are sought from the expert for these areas to be included in the model. Proceeding iteratively, the segmentation is progressively estimated. This highlights that strategically eliciting expert feedback can result in better spatial analyses.