이미지를 강인하게 인식하는 중심주변 조절 기반 뉴럴 네트워크

Abstract

In this study, we propose a novel deep neural network module that is robust to an unstructured environment. Deep neural networks show higher classification performance than human when training and testing data are drawn from similar distribution. However, deep neural networks show significantly lower classification performance than human in corrupted data. In this study, we propose Trainable Surround Modulation that applies the surround modulation function to deep neural networks based on end-to-end learning. Trainable Surround Modulation is trained based on training data and has a surround modulation function so that it can classify the images well for corrupted data. Trainable Surround Modulation is trained on ImageNet200 and the performance of the module is evaluated on the corrupted dataset, ImageNet200-C. Experimental results show that the proposed Trainable Surround Modulation accurately performs image classification for corrupted data.