An Optimized Algorithm for Human Portrait Image Segmentation Using U-Net

Abstract

Segmentation is a technique used in image analysis that involves the division of an image into smaller, more manageable regions corresponding to distinct objects. Image segmentation can be accomplished in a variety of ways from simple hand-specified regions to intelligent auto-detected regions of interest. Regions of interest can be different objects in an image or different color, foreground, and background of an image. Segmentation process is different for each type of application and there is a lack of a universal process that can be applied to all image segmentation tasks. Experts in the field have proposed many Neural Network-based solutions yet unable to achieve significant results segmenting human portraits. To address this issue, this article proposes the use of U-Net model incorporated with alpha matting, for image segmentation of people, separating foreground and background. For experiments, Matting Human Dataset has been used that is publically available on Kaggle. We evaluated the performance of our proposed model and obtained the Jaccard similarity index 0.95 and Dice similarity index 0.72. Empirically, our proposed model takes the advantages of using U-Net model to accomplish reliable results when compared with the other state of the art methods.