Infrared thermal image segmentation using expectation-maximization-based clustering

Abstract

In infrared (IR) based non-destructive and evaluation tests (NDT&E) for automated fault detection and identification processes, the segmentation task is a crucial stage. In fact, thermal imaging gives vital condition information of equipment and structures. So, pattern recognition algorithms can perform an accurate diagnosis, through an adequate segmentation. In this paper the Expectation Maximization Clustering (EM-Clustering) segmentation is evaluated for IR images, using as reference watershed transform-based segmentation. IR images were acquired from a test rig of an operating motor at Vibrations Laboratory. Proposed Clustering based segmentation performance is assessed by Dice's coefficient metric, obtaining an average 0.87 Dice's coefficient value. Demonstrating that EM-Clustering Segmentation is a valid choice for IR image processing.