이미지를 강인하게 인식하는 보조 노이즈 전파 기법

Abstract

In this study, we propose an auxiliary noise propagation module that is robust to an unstructured environment. Deep neural networks show high classification performance when training and testing data are drawn from the independent and identical distribution. However, deep neural networks show significantly low performance for the out-of-distribution data. Augmented noisy images could be the solution, but they degrade the model performance for the original data distribution. Our proposed model balances and improves performance for both cases. Our model propagates the original images to the original batch normalization layer and augmented noisy images to the auxiliary batch normalization layer separately. Separate batch normalization layers enforce the model to learn various features and improve the robustness for out-of-distribution data. Auxiliary noise propagation is trained on CIFAR-10 and the performance of the module is evaluated on the CIFAR-10-C. Experimental results show that the proposed auxiliary noise propagation improves the robustness for out-of-distribution data.