On semantic image segmentation using deep convolutional neural network with shortcuts and easy class extension

Chunlai Wang, Lukas Mauch, Ze Guo, Bin Yang

Abstract: In this paper we examine the use of deep convolutional neural networks for semantic image segmentation, which separates an input image into multiple regions corresponding to predefined object classes. We use an encoder-decoder structure and aim to improve it in convergence speed and segmentation accuracy by adding shortcuts between network layers. Besides, we investigate how to extend an already trained model to other new object classes. We propose a new strategy for class extension with only little training data and class labels. In the experiments we use two street scene datasets to demonstrate the strength of shortcuts, to study the contextual information encoded in the learned model and to show the effectiveness of our class extension method.