【作者】 胡学龙； 高燕； 高和蓓；

【Author】 HU Xue-long1,2+, GAO Yan1, GAO He-bei1 1(Department of Electronics and Communication, Coll of Inf Engin, Yangzhou University, 225009, China) 2(Jiangsu Prov Key Lab for Compt Inf Process Tech, , Suzhou University, Suzhou, 215006, China)

【机构】 扬州大学信息工程学院电子与通信工程系； 苏州大学江苏省计算机信息处理技术重点实验室；

【摘要】 针对基于内容的图像检索问题,本文提出了一种新的基于整数小波分解的彩色图像检索算法。其中颜色特征采用小波低频系数矩阵中环形区域的F-范数,纹理特征则通过高频部分的能量比值来描述。在进行图像间的相似性度量时,为了结合不同的子特征进行全局的相似性检索,采用Guassian模型对不同子特征间的距离作归一化处理。实验结果表明,两种特征的线性结合有效地改进了图像的检索性能,其特征维数少、计算复杂度小的优点使得它十分适合于大型图像库的检索。更多还原

【Abstract】 Color and texture feature vectors of an image are always considered to be an important attribute in content-based image retrieval system. In this paper, a new color image retrieval algorithm using both color feature and texture feature based on the integer-to-integer wavelet transform is presented. First, an image is decomposed for three times using 5/3 reversible integer-to-integer wavelet transform by JPEG2000 standard, most of the energy are concentrated in the low frequency band with only 1/64 size of the original image, which is segmented into different circulars. The circular region F-norms in low frequency band of wavelet transform is used as color feature of the image. This kind of feature vector is invariant to rotation, translation and scale. The similar degree of color feature between two images is measured by a similarity measure similar to S1 similarity measure. Secondly, wavelet transform is very popular in characterizing texture description and classification. Compared with traditional wavelets, 2/10 integer-to-integer wavelet is easier to implement, moreover, the length of low pass filters is shorter, which is 2, and that of high pass filters is longer, which is 10. Many literatures conclude that high frequency bands correspond to texture information, this kind of wavelet that has longer high pass filters is much suitable for extracting texture features of images. An image is decomposed for three times using 2/10 integer-to-integer wavelet transform, nine high frequency bands are derived, average energy of these bands are computed, the ratio of the average energy of diagonal direction and the sum of the average energy of horizontal and the vertical direction in every level is used as texture feature. This kind of feature vector is invariant to rotation 90°,180°and 270°, the similar degree of texture feature between two images is measured by correlation. Every similarity measure has different range, so in matching the similarity of the images, the Guassian model is used to normalize the different sub-characters distance into the same range, the global similarity is the linear integration of color similar degree and texture similar degree. Experiment is done by various image type in an image database, the retrieval performance is evaluated by recall and ranking value, compared with literature [6] which uses color histogram as color feature, and the detail information of the multi-resolution representation of the image as texture feature, our algorithm is more satisfactory. Many experimental results suggest that it is invariant to rotation, translation and scale, immune to illumination. Furthermore, the proposed integration improves retrieval performance, taking full advantage of the rich representation of color and the statistics of the wavelet coefficients, which is with low-dimensional features and fewer computation merits, overcoming the disadvantage of color histogram which losing the spatial information of the color, it’s very suitable for large image database retrieval. In our later research, the user’s factor is taken into account, so we introduce into relevance feedback, in order to satisfy the user’s demands.更多还原