基于多小波统计特征的纹理图像检索

Statistical Characterization of Multiwavelet for Texture Retrieval

为了研究多小波性能,对多小波系数分布的统计特性进行了研究。多小波在实数域能同时具有正交、对称、短紧支撑和高消失矩等特性,单小波却不具有上述的性质,因此在理论上多小波比单小波具有更多的优势。提出并验证了多小波系数直方图服从于指数族分布;根据多小波的特点研究了其系数分布的一阶、二阶矩(共生矩阵)和系数直方图的统计特性,并应用于纹理特征的提取。通过理论分析和在纹理图像检索的对比实验说明在冗余预滤波方式下,采用二阶统计矩方法时多小波优于单小波。

Theoretically, the multiwavelet is better than scalar wavelet, so we analyze the statistical characteristics of the detail wavelet coefficients of multi-wavelet transform, and bring forward the wavelet coefficients histogram of texture image which can be modeled by a family of exponentials. We study the ways to extract the texture features based on the first-order and second-order （co-occurrence） of the statistical characters of multi-wavelet transform. The theoretic analysis and experimental results show that second-order signatures of multi-wavelet are better than that of scalar wavelet using the redundancy prefiltering method.