Deep learning based semantic segmentation using feature augmentation

Abstract

In this paper, we address the problem of semantic image segmentation. The difficulties of training a network interrupt to increase the depth and volume. We address this limitation by augmenting feature from different pre-trained network and improve the performance by incorporating additional information into a semantic segmentation framework. We show that fully-convolutional-network with our feature augmentation technique produces competitive results with state-of-art method in terms of segmentation accuracy but get the upper hand over preserving object boundaries. We believe that straight-forward extension of the proposed approach can be also applied for other recognition task.