Foreground segmentation using multiscale convolutional neural network
Huini Fu, Ben‐Zhang Wang, Hengzhu Liu · Electronics Letters · 2020
Foreground segmentation is to segment moving objects from a video or camera. Many approaches have been proposed to perform background subtraction. However, these algorithms could not handle complex scenes like illumination changes, dynamic background and camouflage etc. In this Letter, the authors propose a robust end‐to‐end convolutional neural network with multiscale features to address these problems. They use only a few training samples with RGB channels and output the corresponding foreground mask, which could save millions in labelling ground‐truth for the big dataset. They evaluate their model using Change Detection 2014 dataset, and results outperformed other state‐of‐the‐art methods with F ‐measure of 0.982 with ∼30 fps processing speed.