Go-selfies: A Fast Selfies Background Removal Method Using ResU-Net Deep Learning

Abstract

The selfies play an important role in recording meaningful moment in human’s daily life. In most cases, before sharing photos, people often synthesis attractive images on some phone applications, such as Photoshop. While these kinds of software have reached good performance nowadays, they are too complex for simple life usage. In this work, we proposed an automatic segmentation model unique to segment human selfies photos. We first constructed a large photo segmentation database and built 8 different models based on resolution, image size and whether or not to use transfer learning and picked the best one among them. We then applied cyclical learning rate method and pre-trained encoder network to fine tune our models. Finally, our best model tested on Google images demonstrated satisfying promising results on both accuracy scores and losses, which will be the precondition in real-time segmentation. We named this lovely web product as Go Selfies .