Semantic Segmentation Based on Deep Convolution Neural Network

作     者：Jichao Shan Xiuzhi Li Songmin Jia Xiangyin Zhang 

作者机构：Faculty of Information Technology Beijing University of Technology Beijing 100124 China Beijing Key Laboratory of Computational Intelligence and Intelligent System Beijing 100124 China. 

出 版 物：《Journal of Physics: Conference Series》 

年 卷 期：2018年第1069卷第1期

学科分类：07[理学] 0702[理学-物理学] 

摘      要：Semantic segmentation using full convolutional neural network (FCN) avoids the problems of repeated calculation and storage due to using of pixel blocks. However, the results obtained by FCN are still not precise enough, and the results of upsampling are still relatively fuzzy and smooth. It is not sensitive to the details such as small object in the image. Therefore, this paper proposes an image segmentation method based on simplified deep residual network. A simplified Deep Residual Network (DRN) is proposed to replace the VGG-16 network in the original FCN framework. The net combines deep residual network architecture with 30% lesser parameters than the VGG model and the method of skip connection that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate and detailed segmentation results. The results showed that the effect of segmentation is more refined and the recognition and segmentation of small objects are significantly improved by the improved network model.