Joint Image Segmentation and Interpretation Using Iterative Semantic Region Growing on SAR Sea Ice Imagery

Abstract

Segmentation of images into disjoint regions and interpretation of the regions for semantic meanings are two central tasks in an image analysis system. Typically, the segmentation and interpretation are performed separately with the interpretation as a post processing of segmentation. In this paper, we use an iterative method that keeps refining the segmentation and producing semantic class labels at the same time. The segmentation algorithm is based on a region growing technique and the interpretation is a Markov Random Field (MRF) based classification. The two processes are integrated under the Bayesian framework, with both aiming at reducing a defined energy. The interactions between the two are bidirectional by letting the interpretation result have some degree of control on the region growing process. Various features can hence be efficiently combined, and accurate classifications are obtained for operational synthetic aperture radar (SAR) sea ice applications.