基于噪声抑制的Fisher向量在图像分类中的应用.docVIP

  基于噪声抑制的 Fisher 向量在图像分类中 的应用 李萌，黄雅平，林杰** 5 10 15 20 25 30 35 40 （北京交通大学计算机与信息技术学院，北京 100044） 摘要：基于 Fisher 向量的图像全局特征表达可以看作是词袋模型（Bag-of-visual-words）的 扩展，二者都是基于图像的一种中间表示。前者由于融合了判别式和生成式模型的优点，不 仅体现了每一个视觉单词的出现频率，还编码了关于局部特征分布的额外信息，因此 Fisher 向量是一种信息量更丰富的图像表示。但是基于传统 Fisher 向量的分类器性能在很大程度 上受限于图像背景或不相关区域中提取到的无意义甚至起干扰作用的局部特征。因此本文提 出一种基于贝叶斯自适应算法的噪声抑制方法，有效去除图像中判别能力差以及严重削弱分 类性能的噪声特征，只将保留下来的局部特征用于 Fisher 向量的计算，实验表明，基于噪 声抑制的 Fisher 向量对于低分辨率图像的分类识别准确率相比传统 Fisher 向量有了显著提 升，并在一定程度上降低了计算复杂度。 关键词：图像分类；Fisher 向量；贝叶斯自适应；噪声抑制；低分辨率图像 中图分类号：TP391.4 Image classification based on Fisher vector with noise suppression LI Meng, HUANG Yaping, LIN Jie (Computer and Information Technology School,Beijing Jiaotong University,Beijing 100044) Abstract: Fisher vector, as a global feature expression, can be seen as an extension of the bag-of-visual-words. Both of them are based on an intermediate representation. But fisher vector is a kind of richer image representation because of gathering the advantages of generative and discriminative approaches. It is not limited to the number of occurrences of each visual word but it also encodes additional information about the distribution of the descriptors. However the classification performance based on traditional fisher vector is largely limited by the uninformative and noisy low-level features from background or irrelevant image segments. We propose an approach of adaptive noise suppression based on Bayesian adaptation. It can effectively remove the poor discriminative and noisy features with negative effect. Then the preserved features are used to compute fisher vector. Our experiment shows that noise suppressed fisher vector (NSFV) perform better on low-resolution, low-quality dataset not only compared to the traditional FV, and can reduce encoder complexity to some extent. Key words: Image classification; Fisher vector; Bayesian adaptation; noise suppression; low-resolution images 0 引言 图像分类识别问题是计算机视觉领域的研究热点，其目的是根据图像的语义内容给其分 配一个或多个标签，由于现实世界图像的多样性以及类内、类