基于迁移鲁棒稀疏编码的图像表示方法

发布时间：2019-03-19 19:04

【摘要】：图像表示是图像处理和图像理解研究中的关键问题之一.在图像的低层表示上有很多重要的研究工作,例如HOG,SIFT等.然而在图像的低层表示和高层语义间仍然存在着巨大的鸿沟.因而,很多机器学习的方法被用来学习图像的高层表示,例如主成分分析,稀疏编码,非负矩阵分解以及低秩表示等.传统机器学习假设标记图像和未标记图像服从同一分布,图像表示的误差服从高斯分布.然而现实中图像数据更新速度快,而且图像生成环境存在差异性,导致未标记图像与已标记图像不服从同一分布,因而需要重新标记数据和训练模型.并且图像数据容易出现异常,例如遮挡、腐蚀等等,从而不能再用高斯分布来估计误差.迁移学习允许标记图像(训练数据)和未标记图像(测试数据)服从不同的分布.基于迁移学习的图像表示方法学习一个新的好的特征空间.在这个新的特征空间下,可以较好地描述标记图像和未标记图像的语义信息.并且在这个新的特征空间下,从训练集中标记图像上学习到的统计模型(例如分类模型),可以较好地迁移到测试集中未标记图像上,从而充分利用已标记图像,将学习到的知识迁移到未标记的图像集上.该文提出了一种基于迁移鲁棒稀疏编码的图像表示方法,引入权值矩阵削弱异常点对分类的干扰,使用稀疏编码获得数据的高级语义,利用最小化最大均值差异缩小源域和目标域图像集之间的分布差异以及图拉普拉斯项保留图像集的几何特性.该文的主要贡献在于:一是通过权值矩阵泛化残差分布,使得所提出的基于迁移鲁棒稀疏编码的图像表示方法能大大减少异常点对编码和字典学习的影响;二是在鲁棒字典学习过程中,采用正则化参数代替迁移稀疏编码中的字典约束,从而将其转化为无约束优化问题,避免了拉格朗日求解法的复杂性.在几个通用迁移学习数据集上的对比实验结果表明,该文所提出的图像表示方法在分类上的平均准确率比其它6种相关主流方法均有不同程度的提高,证明了其有效性和鲁棒性.
[Abstract]:Image representation is one of the key issues in image processing and image understanding. There are many important research work on the low-level representation of images, such as HOG,SIFT. However, there is still a huge gap between the low-level representation and the high-level semantics of the image. Therefore, many machine learning methods are used to study high-level representation of images, such as principal component analysis, sparse coding, non-negative matrix decomposition and low-rank representation. Traditional machine learning assumes that marked image and unmarked image obey the same distribution, and the error of image representation follows Gao Si distribution. However, in reality, the updating speed of image data is fast, and there are differences in the environment of image generation, which results in the disobedience of unmarked image and marked image to the same distribution, so it is necessary to re-label the data and training model. And the image data are easy to appear abnormal, such as occlusion, corrosion and so on, so Gao Si distribution can no longer be used to estimate the error. Migration learning allows marked images (training data) and untagged images (test data) to follow different distributions. Image representation method based on migration learning to learn a new good feature space. In this new feature space, the semantic information of marked image and unmarked image can be well described. And in this new feature space, statistical models (such as classification models) learned from marked images in training sets can be well migrated to untagged images in test sets, thus making full use of tagged images. Migrate the learned knowledge to an untagged set of images. In this paper, an image representation method based on migration robust sparse coding is proposed. The weight matrix is introduced to weaken the interference of outliers on the classification, and the sparse coding is used to obtain the high level semantics of the data. The distribution difference between the source domain and the target domain is reduced by minimizing the maximum mean difference, and the geometric properties of the image set reserved by the Tulapulas term are also reduced. The main contributions of this paper are as follows: firstly, the residual distribution is generalized by weight matrix, so that the proposed image representation method based on migration robust sparse coding can greatly reduce the influence of outliers on coding and dictionary learning; Second, in the process of robust dictionary learning, regularization parameters are used to replace dictionary constraints in migration sparse coding, which is transformed into an unconstrained optimization problem, thus avoiding the complexity of Lagrangian solution. The experimental results on several general transfer learning data sets show that the average classification accuracy of the proposed image representation method is higher than that of the other six mainstream methods to some extent. Its effectiveness and robustness are proved.
【作者单位】：安徽大学计算智能与信号处理教育部重点实验室;安徽大学计算机科学与技术学院;
【基金】：国家自然科学基金(61602004,61472001) 安徽省高校自然科学研究重点项目(KJ2016A041) 安徽省自然科学基金(1408085MF122,1508085MF127) 安徽大学信息保障技术协同创新中心公开招标课题(ADXXBZ2014-5,ADXXBZ2014-6)资助~~
【分类号】：TP391.41

【相似文献】