基于改进群组归一化的目标检测与实例分割

Object Detection and Instance Segmentation Based on Improved Group Normalization

训练卷积神经网络,尤其是在使用ImageNet这种大规模数据集进行深层卷积神经网络训练时,往往需要投入几天甚至几个月的时间。为了加快神经网络的迭代收敛速度,通常需要使用批量归一化算法对训练网络的输入输出分布进行统一调整。本工作结合实例归一化方法对群组归一化进行了改进,并将改进后的该方法结合Faster R-CNN和Mask R-CNN进行了目标检测与实例分割实践。

Training convolutional neural networks,especially deep neural networks with large scale data sets such as ImageNet,usually takes days or even months.In order to accelerate the convergence of the neural network,batch normalization is usually used to normalize the inputs and outputs distribution of the training network.This paper improves group normalization with instance normalization,and combines this method with Faster R-CNN and Mask R-CNN for object detection and instance segmentation.