Open AccessEnd-to-End Deep Residual Network for Semantic Segmentation
Author(s) -
Yingzi Zhou,
Kun Huang,
Xiaoying Guo,
Xiaohai He
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1684/1/012053
Subject(s) - residual , computer science , convolutional neural network , encoder , convergence (economics) , feature (linguistics) , segmentation , dimension (graph theory) , block (permutation group theory) , artificial neural network , artificial intelligence , algorithm , deep learning , activation function , function (biology) , pattern recognition (psychology) , mathematics , linguistics , philosophy , geometry , evolutionary biology , pure mathematics , economics , biology , economic growth , operating system
Recent work has made significant progress in improving for the pixelwise labeling with convolutional neural networks by using residual networks. In this paper, we explore the impact of residual network in semantic segmentation by residual encoder-decoder model. The residual block can improves the dimension of feature maps, and obtains more efficient feature maps, and then ensures the effective of deep network structure. Moreover, we select the activat ion function with zero-centered characteristics to speed up the model convergence. The SELU activation function is used to remove the BN function in the network due to normalized deploy ment, thereby reducing the amount of network calculations. The proposed residual encoder-dec oder model can still output high-precision prediction results without any post-processing, and t he convergence speed is significantly faster. Our approach has achieved 61.13% mIoU, 86.92% PA, 73.42% MPA on Camvid dataset.