eduzhai > Applied Sciences > Engineering >

Adaptive feature recombination and recalibration for semantic segmentation with Fully Convolutional Networks

  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Fully Convolutional Networks have been achieving remarkable results in imagesemantic segmentation, while being efficient. Such efficiency results from thecapability of segmenting several voxels in a single forward pass. So, there isa direct spatial correspondence between a unit in a feature map and the voxelin the same location. In a convolutional layer, the kernel spans over allchannels and extracts information from them. We observe that linearrecombination of feature maps by increasing the number of channels followed bycompression may enhance their discriminative power. Moreover, not all featuremaps have the same relevance for the classes being predicted. In order to learnthe inter-channel relationships and recalibrate the channels to suppress theless relevant ones, Squeeze and Excitation blocks were proposed in the contextof image classification with Convolutional Neural Networks. However, this isnot well adapted for segmentation with Fully Convolutional Networks since theysegment several objects simultaneously, hence a feature map may containrelevant information only in some locations. In this paper, we proposerecombination of features and a spatially adaptive recalibration block that isadapted for semantic segmentation with Fully Convolutional Networks - the SegSEblock. Feature maps are recalibrated by considering the cross-channelinformation together with spatial relevance. Experimental results indicate thatRecombination and Recalibration improve the results of a competitive baseline,and generalize across three different problems: brain tumor segmentation,stroke penumbra estimation, and ischemic stroke lesion outcome prediction. Theobtained results are competitive or outperform the state of the art in thethree applications.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...