Abstract: Aiming at spectrum aliasing problems in the directional subbands during the contourlet transform, considering the limitation of the traditional measure KL Distance(KLD) between two hidden Markov models, this paper proposes a texture image retrieval method based on improved KLD, using Non-Aliasing Contourlet Transform(NACT) Hidden Markov Tree(HMT) model. The algorithm uses NACT to decompose a texture, which can deal with the spectrum aliasing phenomenon well, trains the HMT model and takes the HMT model parameter set as the texture features. It computes the similarity between two models using improved KLD, which meets the triangle inequality properties and can measure the distance better. The proposed algorithm is verified by theory and experiment, and results show that the precision of proposed method improves 2.81 percent than that of the CT-HMT combining traditional KLD measure method.

Key words: contourlet, Non-Aliasing Contourlet Transform(NACT), Hidden Markov Tree(HMT), improved KLD method

摘要： 针对轮廓波变换方向子带中的频谱混叠现象及传统KLD方法度量隐马尔科夫模型间距离的局限性，提出结合改进KLD度量的抗混叠轮廓波隐马尔科夫树(HMT)纹理图像检索方法。利用抗混叠轮廓波变换抑制频谱混叠的特点对纹理进行分解，建立HMT模型并将其训练后的参数集视为纹理特征，利用改进KLD方法满足三角不等式的优点度量HMT模型间的距离，提高纹理图像检索精度。理论和实验结果表明，该算法的查准率比CT-HMT+传统KLD方法提高了2.81%。

关键词: 轮廓波, 抗混叠轮廓波变换, 隐马尔科夫树, 改进KLD方法

CLC Number: 

• TP391.4