旋转不变纹理特征用于两级图像检索

针对图像中常见的旋转问题提出一种旋转不变纹理特征进行两级图像检索的方法。粗检中,通过坐标变换把图像的旋转转换为行移,并提取近似行移不变的小波特征,结合粗比较算法对整个图像库进行粗检。然后对通过粗检的图像进行 Gabor 变换,提取旋转不变精检索特征,并使用Canberra 距离进行相似性度量。通过对旋转图像库的测试表明,该方法不仅加快了运算速度,且当参数选择适当时,在相同特征条件下,检索率比直接使用精检索方法检索时还提高了 1.625%。

Rotation-invariant texture feature for two-stage image retrieval

A method used for two-stage image retrieval with rotation-invariant texture feature is proposed against the common rotation problem of image. In coarse retrieval, the image is converted into a rotation-invariant but row-shifted image through polar transform. The approximate row-shifted invariant wavelet feature is extracted. Combined with coarse comparison algorithm, a coarse retrieval will be carried out for the whole image database to remove undesired images. The image after coarse retrieval will be processed through Gabor transform and a rotation-invariant feature will be got to perform the fine matching. The similarity is measured by Canberra distance. The results show that the method speeds up operation speed. When suitable parameters are selected under conditions of the same feature, its retrieval ratio has been improved for 1.625% higher than that of the fine retrieval method.