图像数据集噪声对卷积网络分类的影响
首发时间:2018-03-09
摘要:在图像分类任务中准备数据集时,数据噪声的干扰会对后续的分类模型产生不利影响,噪声越严重就越难以得到合理的分类模型。为了评价数据集噪声的影响,本文研究了多个不同噪声类型数据对深度卷积模型的干扰。论文基于Cifar-10基准数据集设置了多种数据噪声类型,然后利用这些含噪声的数据训练深度卷积模型进行分类,实验结果表明:数据集之中噪声会对深度卷积网络分类模型带来明显的不利影响,其中随机噪声产生的不利影响较小,但是类别之间的交叉噪声图像会显著降低模型的识别能力,已训练模型对初始训练数据进行筛选可以有效避免交叉噪声污染。
For information in English, please click here
Influence of Image Dataset Noise on Classification of Convolutional Network
Abstract:In data set preparation stage for image classification task, the noise of data set will adversely affect the classification model, and the more serious the noise, the more difficult it is to get a reasonable classification model. In order to evaluate the effect of data set noise, this paper studies the harmful effect of several different types of noise on the deep convolution model. Based on Cifar-10, this paper sets up a variety of types of noise, and then use these data containing noise to training the deep convolutional model. Experimental results show that the data set noise can lead to obviously adverse effect on convolutional network, in which the effect of random noise is slight, but the crossing noise can significantly weaken the recognition ability, and that to filter the initial training data with trained model can effectively avoid crossing noise pollution.
Keywords: image recognition dataset noise deep convolutional network
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
图像数据集噪声对卷积网络分类的影响
评论
全部评论0/1000