Self-Supervised Learning for Segmentation using Image Reconstruction
Srivallabha Karnam · RIT Scholar Works (Rochester Institute of Technology) · 2020
Deep learning is the engine that is piloting tremendous growth in various segments of the industry by consuming valuable fuel called data. We are witnessing many businesses adopting this technology be it healthcare, transportation, defense, semiconductor, or retail. But most of the accomplishments that we see now rely on supervised learning. Supervised learning needs a substantial volume of labeled data which are usually annotated by humans- an arduous and expensive task often leading to datasets that are insufficient in size or human labeling errors. The performance of deep learning models is only as good as the data. Self-supervised learning minimizes the need for labeled data as it extracts the pertinent context and inherited data content. We are inspired by image interpolation where we resize an image from a one-pixel grid to another. We introduce a novel self-supervised learning method specialized for semantic segmentation tasks. We use Image reconstruction as a pre-text task where pixels and or pixel channel (R or G or B pixel channel) in the input images are dropped in a defined or random manner and the original image serves as ground truth. We use the ImageNet dataset for a pretext learning task, and PASCAL V0C to evaluate efficacy of proposed methods. In segmentation tasks decoder is equally important as the encoder, since our proposed method learns both the encoder and decoder as a part of a pretext task, our method outperforms existing self-supervised segmentation methods.