Scene classification via learning a multi-branch convolutional network
Department of Applied Computing
Efficient and accurate classification of high-resolution scene remains a challenge of within-class diversity and between-class similarity due to rich image variations in viewpoint, object pose, spatial resolution and background. To address these issues, we propose a multi-branch convolutional neural network (MB-CNN), which focuses on tackling the problem of learning the appropriate representation of a high-resolution scene that is rich enough to discriminate between different semantic classes. First, pyramid scene parsing network (PSPNet) with minor modification is introduced to gather global object information. Then, an attention net is proposed to highlight transformation invariance and key regions for attention feature extraction. Finally, above two branches are fused with original input branch to learn consistently semantic class information and generate powerful predictions. Our approach achieves better performance favorably against state-of-the-arts on two publicly available scene datasets.
2019 IEEE International Conference on Systems, Man and Cybernetics
Scene classification via learning a multi-branch convolutional network.
2019 IEEE International Conference on Systems, Man and Cybernetics.
Retrieved from: https://digitalcommons.mtu.edu/michigantech-p/1326