Segmentation of Portrait Images Using A Deep Residual Network Architecture

Özet

Segmenting portrait images into semantic areas is an important step towards scene understanding and image analysis. Although segmentation is a very active field of study, there are few studies in the field of portrait segmentation. One of the most crucialsteps in portrait segmentation is the precise segmentation process where semantically related pixels grouped together including hair, face, body, and background. However, this is a challenging problem due to the extreme variations in hair shape, color, and background. In order to handle such variations, we proposed a deep residual network based on ERFNet architecture. We used geometrically normalized faces as an input for the network. Experimental studies on EG1800 dataset (two-classes) and LFW Part LabelsDataset (three-classes) showed that the proposed method provides state of the art mIoU (mean intersection over union) and pixel-based accuracy. We obtained 96.37% mIoU and 98.17% pixel based accuracy for EG1800 dataset and 90.1% mIoU and 97.14% accuracy for the LFW dataset.