Method and device for learning fog-invariant feature
AssigneePostech Research and Business Development Foundation
InventorsSuha KWAK, Taeyoung SON, Sohyun LEE
Life of this patent
Estimating the term…
Abstract
A device for learning a fog-invariant feature includes a fog-pass filter, which is trained such that fog factors of the same fog domain approach each other and fog factors of different fog domains are apart from each other so as to determine a fog state of an input image as a fog factor, and a segmentation network, which is trained such that a fog style gap between the different fog domains decreases, in which training of the fog-pass filter and training of the segmentation network are alternatively performed by using three fog domain datasets.
Claims15
A device for learning a fog-invariant feature, the device comprising:
- a fog-pass filter, which is trained such that fog factors of a same fog domain approach each other and fog factors of different fog domains are apart from each other so as to determine a fog condition of an input image as a fog factor;
- and
- a segmentation network, which is trained such that a fog style gap between the different fog domains decreases,
- wherein training of the fog-pass filter and training of the segmentation network are alternatively performed by using three fog domain datasets.
A method of learning a fog-invariant feature, the method comprising:
- representing an effect of fog on an input image as a fog factor in a fog-pass filter connected to a segmentation network and receiving, as an input, a Gram matrix of a feature map of the segmentation network;
- determining a fog condition of the input image as a fog factor by fixing a parameter of the segmentation network and training, in the fog-pass filter, to draw fog factors of a same fog domain together and hold fog factors of different fog domains apart;
- and
- fixing the parameter of the fog-pass filter, and training, in the segmentation network, to close a fog style gap between different fog domains,
- wherein training of the fog-pass filter and training of the segmentation network are alternatively performed.
Dependencies are inferred from the claim text.