A Study on the Improvement Performance of Object Segmentation using Group Normalization method

GeunTaeKim, Yeonghun Lee,HyungHwa Ko,Kyounghak Lee

Recently, divers Image Processing technics works in our daily life. One of them is Object Segmentation to find region of interest in images and to classify. To apply this system on real word, faster speed is essential. Therefore we propose method to enhance performance of Segmentation by varying Normalization.Related works for Object Segmentation are DeepLab V3+, BiSeNet, Mask R-CNN and so on. We propose method to enhance performance of each model by applying Group Norm but not Mask R-CNN, because it uses Instance segmentation. The model that batch size is lower than 8 outperforms when Group Normalization is applied. Further Training time relatively is decreased a lot. The model with BiSeNet favorably works in real-time but not model with Deeplab V3+The use of Group Normalization reduced training time and improved performance.

Volume 11 | 05-Special Issue

Pages: 2439-2445