一种图像场景的独立子空间ISA分类方法

Independent Subspace Analysis Classification Method for Image Scene

随着计算机应用和多媒体的不断发展与应用,数字图像变得越来越多,内容越来越丰富.如何在海量的图像数据中获得想要的和有用的信息也变得越来越重要.图像场景分类就是其中一种重要的技术.本文采用了基于独立子空间分析(ISA)网络模型的特征提取方法并结合空间金字塔匹配(SPM)模型和支持向量机(SVM)分类器实现对图像场景的分类.基于ISA(独立子空间分析)网络模型的特征提取方法是一种无监督学习方法,能够获取图像中结构化的特征基元,并在规则网格划分的策略下利用所得的结构化的特征基元获取图像块描述子.然后结合空间金字塔匹配(SPM)模型构建金字塔结构式的整幅图像特征表示.实验在Scene-15图像场景数据集的基础上进行,并将本文方法与基于尺度不变特征转换(SIFT)特征提取方法的几种常用经典方法进行对比实验,实验结果表明本文方法在选取了合适的特征基元个数后,提高了提取图像特征的速度和时间以及图像场景的分类准确率.

With the development and application of computer application and multimedia,the number of digital images is getting more and more,the content is getting richer. How to get the desired and useful information in massive image data is becoming more and more important. Image scene classification is one of the important technology. In this paper,a feature extraction method based on Independent Subspace Analysis（ ISA） network model is adopted and the classification of image scene is realized by combining Spatial Pyramid Matching（ SPM） model and Support Vector Machine（ SVM） classifier. The feature extraction method based on Independent Subspace Analysis（ ISA） network model is an unsupervised learning method,which can obtain the structural feature bases in the image and the patch-based image descriptors are computed over regularly divided using the structured feature bases. Then,these descriptors are taken into the Spatial Pyramid Matching（ SPM） model,which incorporates spatial layout information and global geometric correspondence for recognizing image scene categories,to build whole image feature representations. Compared with several commonly used classical methods based on Scale-invariant feature transform（ SIFT） feature extraction in classification task on Scene-15 image scene data set,the proposed method improves the speed and time of extracting image features and the classification accuracy of image scenes after selecting the appropriate number of feature bases.