Journal of Frontiers of Computer Science and Technology ›› 2019, Vol. 13 ›› Issue (12): 2138-2148.DOI: 10.3778/j.issn.1673-9418.1809054

Previous Articles     Next Articles

Multiple Representations for Image Classification Approaches

CHEN Deyun, FU Lijun, ZHANG Xuesong, YU Liang, CHEN Hailong, LI Ao   

  1. 1.School of Computer Science and Technology, Harbin University of Science and Technology, Harbin 150080, China
    2.Beijing Zhaoxin Semiconductor Co., Ltd., Beijing 100084, China
    3.Jiuquan Satellite Launch Center, Dunhuang, Gansu 736200, China
  • Online:2019-12-01 Published:2019-12-10

多种表示的图像分类方法

陈德运付立军张学松于梁陈海龙李骜   

  1. 1.哈尔滨理工大学 计算机科学与技术学院,哈尔滨 150080
    2.北京兆芯集成电路有限公司,北京 100084
    3.酒泉卫星发射中心,甘肃 敦煌 736200

Abstract: Extracted salient features from different images are significant for image classification. However, single image classification methods are not robust for all the images. This paper extracts different features by multiple views and fuses them to effectively address the above problem. Firstly, this paper uses 2DPCA (two-dimensional principal component analysis) to extract features of images, which are utilized to reconstruct images (also referred to as virtual images). Secondly, it applies FFT (fast Fourier transform) to obtain spectrum features. Then, it exploits sparse method to get scores from original images, virtual images and spectrum features, respectively. Finally, this paper uses a novel fusion mechanism to integrate the obtained scores above, and classifies the images according to the newly acquired scores. The obtained multiple features are complementary with original images, which makes the proposed algorithm more robust. The proposed method is sparse, which improves the performance of image classification. In addition, it can automatically extract parameters rather than manual settings. Experimental results show that this algorithm has higher accuracy of image classification under different situations.

Key words: image classification, two dimension principal component analysis (2DPCA), fast Fourier transform (FFT), sparse method, multiple representation methods

摘要: 针对不同图像提取出显著的特征对于图像分类是非常有意义的,然而单一图像分类方法不能在所有图像上都取得好的鲁棒性,通过多视角来提取不同特征,并将其融合来有效地解决这个难题。首先,利用二维主成分分析(2DPCA)提取图像的特征,然后根据获取特征进行图像重构(虚拟图像)。其次,利用快速傅里叶变换(FFT)获取图像的频谱特征。接着,将原始图像、虚拟图像、频谱特征分别利用稀疏方法获取得分。最后,利用一种新颖融合机制将上述得分进行融合,并根据新获取得分进行图像分类。获取的多特征和原始图像进行了互补,使该算法更具有鲁棒性;该方法具有稀疏性,提高了图像分类的性能;此外,它能自动获取参数,不需要手动调参。实验结果表明,该方法在不同情景下具有高的图像分类准确率。

关键词: 图像分类, 二维主成分分析(2DPCA), 快速傅里叶变换(FFT), 稀疏方法, 多种表示方法