基于数据均衡的增进式深度自动图像标注

发布时间：2018-03-17 12:38

本文选题：SAE(stacked　切入点：auto-encoder)　出处：《软件学报》2017年07期 　论文类型：期刊论文

【摘要】：自动图像标注是一个包含众多标签、多样特征的富有挑战性的研究问题,是新一代图像检索与图像理解的关键步骤.针对传统的基于浅层机器学习标注算法标注效率低下、难以处理复杂分类任务的问题,提出了基于栈式自动编码器(stacked auto-encoder,简称SAE)的自动图像标注算法,提升了标注效率和标注效果.主要针对图像标注数据不平衡问题,提出两种解决思路:对于标注模型,提出一种增强训练中低频标签的平衡栈式自动编码器(B-SAE),较好地改善了中低频标签的标注效果.并在该模型的基础上提出一种分组强化训练B-SAE子模型的鲁棒平衡栈式自动编码器算法(RB-SAE),提升了标注的稳定性,从而保证模型本身具有较强的处理不平衡数据的能力;对于标注过程,以未知图像作为出发点,首先构造未知图像的局部均衡数据集,并判定该图像的高低频属性以决定不同的标注过程,局部语义传播算法(SP)标注中低频图像,RB-SAE算法标注高频图像,形成属性判别的标注框架(ADA),保证了标注过程具有较强的应对不平衡数据的能力,从而提升整体图像标注效果.通过在3个公共数据集上进行实验验证,结果表明,该方法在许多指标上相比以往方法均有较大提高.
[Abstract]:Automatic image tagging is a challenging research problem with many tags and various features, which is the key step of the new generation image retrieval and image understanding. It is difficult to deal with the problem of complex classification tasks. This paper proposes an automatic image tagging algorithm based on stack automatic encoder stacked auto-code (SAE), which improves the efficiency and effect of image tagging. Two solutions are proposed: for the annotation model, In this paper, a balanced stack automatic encoder for enhancing low frequency tags in training is presented, which can improve the labeling effect of middle and low frequency tags. Based on this model, a robust balance stack for grouping reinforcement training B-SAE submodel is proposed. The RB-SAE algorithm improves the stability of tagging. In order to ensure that the model itself has a strong ability to deal with unbalanced data, for the annotation process, the unknown image as the starting point, first of all, construct the unknown image of the local equilibrium data set. The high and low frequency attributes of the image are determined to determine the different annotation processes. In the local semantic propagation algorithm (SPV), the low frequency image is labeled by the RB-SAE algorithm, and the high frequency image is labeled by the RB-SAE algorithm. The tagging framework for attribute discrimination can ensure that the labeling process has a strong ability to deal with unbalanced data, thus improving the overall image tagging effect. The experimental results on three common data sets show that, Compared with the previous methods, this method has a great improvement in many indexes.
【作者单位】：福州大学数学与计算机科学学院;福建省网络计算与智能信息处理重点实验室(福州大学);
【基金】：国家自然科学基金(61502105) 福建省科技引导性项目(2017H0015) 福建省中青年教师教育科研项目(JA15075)~~
【分类号】：TP391.41

【相似文献】