High Accuracy Video Foreground Segmentation Based on Feature Normalization
Yuma Suzuki, Koichi Ichige · 2021
We propose an efficient foreground segmentation method for video images. The methods using Convolutional Neural Networks (CNNs) have shown high accuracy in the field of video processing. Batch normalization (BN) is a compensation method for CNNs, which controls the weights so as to converge in scale when the learning rate is increased. Therefore it improves the learning speed and accuracy. However, there is a problem that the accuracy decreases when the batch size is small, and it is not optimal for video images where the batch size cannot be increased. In order to solve this problem, a normalization method called Group Normalization (GN), which divides channels into groups and normalizes them, has been attracting attention. In this paper, we improve the accuracy of foreground segmentation of video images by introducing GN. We also propose Layer Instance Normalization (LIN), which is a combination of the two normalization methods. The effectiveness of the proposed methods is demonstrated through simulation.