Indoor Scene Semantic Segmentation Based on RGB-D Image and Convolution Neural Network

Abstract

In recent years, convolutional neural network has been widely used in image semantic segmentation and achieved great success. In this paper, a semantic segmentation network of indoor scene based on rgb-d image is proposed: SRNET (Strong supervision Residual Net). In this network model, the original data is processed by separate training and gradual fusion, and the mandatory supervision module is added in the decoding stage, which effectively improves the accuracy of semantic segmentation. At the same time, anti residual decoding method and jump structure are introduced to reduce information loss. Experimental results show that the segmentation accuracy of this model is better than most of the current segmentation algorithms.