CGNet:cross-guidance network for semantic segmentation  

作　　者：Zhijie ZHANG Yanwei PANG 

机构地区：[1]School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China

出　　处：《Science China(Information Sciences)》2020年第2期49-64,共16页中国科学（信息科学）（英文版）

基　　金：supported in part by the Science and Technology Innovation 2030-Major Project of Artificial Intelligence of the Ministry of Science and Technology of China(Grant No.2018AAA01028);part by National Natural Science Foundation of China(Grant No.61632018).

摘　　要：Semantic segmentation is a fundamental task in image analysis.The issue of semantic segmentation is to extract discriminative features for distinguishing different objects and recognizing hard examples.However,most existing methods have limitations on resolving this problem.To tackle this problem,we identify the contributions of the edge and saliency information for segmentation and present a novel end-to-end network,termed cross-guidance network(CGNet)to leverage them to benefit the semantic segmentation.The edge and saliency detection network are unified into the CGNet,and model the intrinsic information among them,guiding the process of extracting discriminative features.Specifically,the CGNet attempts to extract segmentation,edge,and salient features,simultaneously.Then it transfers them into the cross-guidance module(CGM)to generate the pre-knowledge features based on the modeled information,optimizing the context feature extraction process.The proposed approach is extensively evaluated on PASCAL VOC 2012,PASCAL-Person-Part,and Cityscapes,and achieves state-of-the-art performance,demonstrating the superiority of the proposed approach.Semantic segmentation is a fundamental task in image analysis. The issue of semantic segmentation is to extract discriminative features for distinguishing different objects and recognizing hard examples.However, most existing methods have limitations on resolving this problem. To tackle this problem, we identify the contributions of the edge and saliency information for segmentation and present a novel end-to-end network, termed cross-guidance network(CGNet) to leverage them to benefit the semantic segmentation. The edge and saliency detection network are unified into the CGNet, and model the intrinsic information among them, guiding the process of extracting discriminative features. Specifically, the CGNet attempts to extract segmentation, edge, and salient features, simultaneously. Then it transfers them into the cross-guidance module(CGM) to generate the pre-knowledge features based on the modeled information, optimizing the context feature extraction process. The proposed approach is extensively evaluated on PASCAL VOC 2012,PASCAL-Person-Part, and Cityscapes, and achieves state-of-the-art performance, demonstrating the superiority of the proposed approach.

关 键 词：SEMANTIC SEGMENTATION fully convolutional networks pyramid NETWORK edge DETECTION saliency DETECTION cross-guidance 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]