Log-Gabor系数相关结构及其图像检索应用

Dependency Structural of Log-Gabor Coefficients For Image Retrieval

自然图像频谱具有非高斯性的特点,而Gabor滤波器频谱具有高斯特征,滤波后不能得到自然图像分布的完整频率信息。Log-Gabor函数的传递函数在高频端有一个延长的尾巴,适合对频谱具有厚拖尾的非高斯频谱图像进行建模。对不同方向和尺度下Log-Gabor变换系数采用广义高斯模型(Generized Gaussian Model,GGM)进行建模,模型参数采用期望最大(Expectation Maximization,EM)进行估计。每张图片经过Log-Gabor变换和广义高斯建模处理得到一个多个广义高斯概率密度函数,图像之间的匹配转化为广义高斯概率密度函数之间的距离匹配,而广义高斯概率密度函数之间的距离采用Kullback-Leible距离(K-L距离)函数来定义。在二个纹理数据库和一个人脸数据库上进行了累加匹配特性曲线(cumulative match characteristic,CMC)和正确识别率相关实验,识别率显示Log-Gabor系数相关结构算法比几种相关算法识别率要高。试验充分证实了基于Log-Gabor和广义高斯模型的纹理检索方法的有效性。

The natural image spectrum is non-Gaussian.However,the Gabor filter spectrum is Gaussian which makes it difficult to obtain the whole spectrum information of natural images after filtering.The transfer function of the Log-Gabor function has an extended tail at the high-frequency end,which is suitable for modeling natural images with thick-tailed and non-Gaussian spectrums.The generalized Gaussian model(GGM)is used for modeling Log-Gabor transform coefficients under different directions and scales,and the model parameters are estimated by expectation maximization(EM).Each image is processed by Log-Gabor transform and generalized Gaussian modeling to obtain several generalized Gaussian probability density functions.The matching between images is transformed into the distance matching between generalized Gaussian probability density functions.The distance between generalized Gaussian probability density functions is defined by the Kullback-Leible distance(K-L distance)function.Experiments on two texture databases and one face database have been done about cumulative match characteristic(CMC)and recognition rate.The recognition rate is superior to other related works.Experiments have fully demonstrated the effectiveness of the texture retrieval method based on Log-Gabor and the generalized Gaussian model.