Jaccard Index in Ensemble Image Segmentation: An Approach
Daniel Ogwok, Elizabeth M. Ehlers
Abstract
Many methods have been applied to image segmentation, including unsupervised, supervised, and even deep learning-based models. Semantic and instance segmentation are the two most widely researched forms of segmentation. It is of value to use multiple methods to segment an image. In this paper, we present an image segmentation ensemble methodology. Multiple image segmentation methods are applied to an image and merged to create one segmentation using the proposed method. The technique uses the Jaccard index algorithm, sometimes called the Jaccard similarity coefficient and commonly known as Intersection over Union (IoU). This resulted in better segmentation results than the respective individual segmentation methods. This experiment was applied to mathematical expression recognition (MER), with the expressions taken from blackboards with varying degrees of noise, and lighting conditions, from different classroom environments. A summary of empirical results from the segmentation of multiple images is presented in the paper.