基于自相关图像的纹理特征检索的研究

RESEARCH ON TEXTURE FEATURE RETRIEVAL BASED ON AUTOCORRELATION IMAGES

针对图像检索时的平移、旋转及尺度变化问题,提出了一种基于自相关图像的不变性纹理特征提取方法。首先,用FFT、IFFT快速算法计算图像的自相关图像,消除平移影响,然后对自相关图像进行log-极坐标变化,这样就将旋转和尺度变化转为了平移,再用具有平移不变性的双树复小波进行分解,就可以提取出平移、旋转和尺度不变的特征向量。采用Canberra距离进行相似性度量。通过对发生几何变化的纹理图像库的实验表明,该方法对图像的平移、旋转和尺度变化具有较好的鲁棒性。

Considering the translation,rotation and scale distortion problem during image retrieval, a novel texture feature extraction tech- nique is proposed. Firstly,the FFT and IFFT algorithms are used to calculate the autocorrelation image. Then, log-polar transformation of the autocorrelation image is applied to convert the geometric distorted image into translation, rotation and scale invariant but row-shifted image. Du- al-tree complex wavelet transformation is introduced to extract the invariant features. The similarity is measured by Canberra distance. The experiments carried on translation,rotation and scale database suggest that the proposed approach is insensitive to geometric distortions,