A novel reformed normaliser free network with U-Net architecture for semantic segmentation

Abstract

Recently developed semantic segmentation network architectures include BatchNorm layer and skip connections. They are outperforming with latest training techniques, but the BatchNorm has implicit limitations such as gradients calculation and memory overhead. Hence this paper proposes a novel architecture named as NF-Unet, that combines the simple, flexible and general framework of NF-Nets and the unique architecture of encoder decoder format of U-Net network that can train with huge batch sizes. The backbone of the contracting path consists of NF-net UNet for encoding the image, for identifying the objects in the image. The proposed architecture achieved 87.37 and 70.12 mean intersection over union (mIoU) on train and test dataset and outperforms the other approaches in the literature in terms of number of parameters.