Real-Time Image Semantic Segmentation Networks with Residual Depth-Wise Separable Blocks

Abstract

Semantic image segmentation plays a key role in obtaining pixel-level understanding of images. In recent years, researchers have tackled this problem by using deep learning methods instead of traditional computer vision methods. Because of the development of technologies like autonomous vehicles and indoor robots, segmentation techniques, that have not only high accuracy but also the capability of running in real-time on embedded platform and mobile devices, are in high demand. In this work, we have proposed a new convolutional module, named Residual depth-wise separable, and a fast and efficient convolutional neural network for segmentation. The proposed method is compared against other state of the art real-time models. The experiment results illustrate that our method is efficient in computation while achieves state of the art performance in term of accuracy.